Skip to content

[FEATURE SUPPORT] Move scaling out of streaming loops, bias-initialized acc_s, and fix dQ double-scaling #203

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 11 commits into from
Nov 4, 2025
Merged

[FEATURE SUPPORT] Move scaling out of streaming loops, bias-initialized acc_s, and fix dQ double-scaling #203

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
b6d2d5a
Scale Q by softmax factor before streaming K/V; add sync for smem path
LoserCheems Nov 3, 2025
7840038
Simplifies mask application
LoserCheems Nov 3, 2025
ce06d5f
Streamlines attention init
LoserCheems Nov 3, 2025
2e87f4a
Improves attn block setup
LoserCheems Nov 3, 2025
deffed7
Adds softmax unscale parameter
LoserCheems Nov 3, 2025
c65c642
Sets inverse softmax scaling
LoserCheems Nov 3, 2025
df74af5
Simplifies mask application logic
LoserCheems Nov 3, 2025
fefb7a9
Synchronizes async preload before Q scaling
LoserCheems Nov 4, 2025
f3102e1
Stops redundant dQ scaling
LoserCheems Nov 4, 2025
77c9e80
Refines bias/mask prep in flash bwd
LoserCheems Nov 4, 2025
2c35c89
Drops redundant softmax unscale
LoserCheems Nov 4, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
87 changes: 31 additions & 56 deletions csrc/flash_dmattn/src/flash_bwd_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -644,17 +644,37 @@ inline __device__ void compute_dq_dk_dv_1colblock(const Params &params, const in
clear(acc_dv);
if constexpr (Has_bias) { if (accum_dbias) { clear(acc_dbias); } }

cute::cp_async_wait<0>();
__syncthreads();

// Scale K once before streaming loop Q
#pragma unroll
for (int k = 0; k < size(tKsK); ++k) {
tKsK(k) = static_cast<Element>(tKsK(k) * params.scale_softmax);
}

for (; m_block >= m_block_min; --m_block) {
Tensor acc_s = partition_fragment_C(tiled_mma_sdp, Shape<Int<kBlockM>, Int<kBlockN>>{}); // (MMA=4, MMA_M, MMA_N)
cute::cp_async_wait<0>();
__syncthreads();

Tensor acc_dp = partition_fragment_C(tiled_mma_sdp, Shape<Int<kBlockM>, Int<kBlockN>>{}); // (MMA=4, MMA_M, MMA_N)
Tensor acc_dq = partition_fragment_C(tiled_mma_dq, Shape<Int<kBlockM>, Int<kHeadDim>>{}); // (MMA=4, MMA_M, MMA_K)
cute::cp_async_wait<0>();
__syncthreads();

if (any_active) {
clear(acc_s);
if constexpr (Has_bias) {
// Copy bias from smem to registers
Tensor tSrBias = make_tensor<Element>(shape(acc_s));
Tensor tSrBias_copy_view = smem_thr_copy_PdS.retile_D(tSrBias);
cute::copy(smem_tiled_copy_PdS, tSsBias, tSrBias_copy_view);
#pragma unroll
for (int i = 0; i < size(acc_s); ++i) { acc_s(i) = tSrBias(i); }
} else {
clear(acc_s);
}
}


Comment on lines +675 to 676
Copy link

Copilot AI Nov 4, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Empty line with trailing whitespace. Remove the trailing whitespace for consistency.

Suggested change

Copilot uses AI. Check for mistakes.
if (any_active) {
Tensor dP_sum = make_fragment_like(lse);
#pragma unroll
for (int mi = 0; mi < size(lse); ++mi) { dP_sum(mi) = gdPsum(get<0>(taccScS_row(mi))); }
Expand Down Expand Up @@ -686,71 +706,26 @@ inline __device__ void compute_dq_dk_dv_1colblock(const Params &params, const in
FLASH_NAMESPACE::calculate_dtanh(scores, dtanh, params.softcap);
}

if constexpr (Has_mask && Has_bias) {
// Copy mask and bias from smem to registers
Tensor tSrMask = make_tensor<Element>(shape(acc_s));
Tensor tSrMask_copy_view = smem_thr_copy_PdS.retile_D(tSrMask);
cute::copy(smem_tiled_copy_PdS, tSsMask, tSrMask_copy_view);
Tensor tSrBias = make_tensor<Element>(shape(acc_s));
Tensor tSrBias_copy_view = smem_thr_copy_PdS.retile_D(tSrBias);
cute::copy(smem_tiled_copy_PdS, tSsBias, tSrBias_copy_view);

// Reshape mask, bias from (MMA=4, MMA_M, MMA_N) to (row=(2, MMA_M), col=(2, MMA_N))
Tensor mask = make_tensor(tSrMask.data(), FLASH_NAMESPACE::convert_layout_acc_rowcol(tSrMask.layout()));
Tensor bias = make_tensor(tSrBias.data(), FLASH_NAMESPACE::convert_layout_acc_rowcol(tSrBias.layout()));

// TD [2023-07-29]: I was thinking that we don't need to mask out the elements beyond
// actual_seqlen_k, because acc_s would be some finite value for those indices.
// In the end when we multiply with K to get dQ, the corresponding values of K would be 0,
// so the result would still be correct.
// However, it's possible that the values in acc_s are so large that they overflow
// when we multiply with dP and convert to fp16, resulting in Inf in dS and NaNs in dQ.
// So we need to mask out the elements beyond actual_seqlen_k.
FLASH_NAMESPACE::apply_mask</*Causal_mask=*/Is_causal, /*Has_mask=*/Has_mask, /*Has_bias=*/Has_bias>(
scores, mask, bias, params.scale_softmax,
n_block * kBlockN + (tidx / 32 / AtomLayoutMS) * MMA_N_SdP * 16,
binfo.actual_seqlen_k,
m_block * kBlockM + get<0>(taccScS_row(0)),
binfo.actual_seqlen_q,
AtomLayoutMS * 16
);
} else if constexpr (Has_mask && !Has_bias) {
if constexpr (Has_mask) {
// Copy mask from smem to registers
Tensor tSrMask = make_tensor<Element>(shape(acc_s));
Tensor tSrMask_copy_view = smem_thr_copy_PdS.retile_D(tSrMask);
cute::copy(smem_tiled_copy_PdS, tSsMask, tSrMask_copy_view);

// Reshape mask from (MMA=4, MMA_M, MMA_N) to (row=(2, MMA_M), col=(2, MMA_N))
Tensor mask = make_tensor(tSrMask.data(), FLASH_NAMESPACE::convert_layout_acc_rowcol(tSrMask.layout()));

FLASH_NAMESPACE::apply_mask</*Causal_mask=*/Is_causal, /*Has_mask=*/Has_mask, /*Has_bias=*/Has_bias>(
scores, mask, /*bias=*/nullptr, params.scale_softmax,
n_block * kBlockN + (tidx / 32 / AtomLayoutMS) * MMA_N_SdP * 16,
binfo.actual_seqlen_k,
m_block * kBlockM + get<0>(taccScS_row(0)),
binfo.actual_seqlen_q,
AtomLayoutMS * 16
);
} else if constexpr (!Has_mask && Has_bias) {
// Copy bias from smem to registers
Tensor tSrBias = make_tensor<Element>(shape(acc_s));
Tensor tSrBias_copy_view = smem_thr_copy_PdS.retile_D(tSrBias);
cute::copy(smem_tiled_copy_PdS, tSsBias, tSrBias_copy_view);

// Reshape bias from (MMA=4, MMA_M, MMA_N) to (row=(2, MMA_M), col=(2, MMA_N))
Tensor bias = make_tensor(tSrBias.data(), FLASH_NAMESPACE::convert_layout_acc_rowcol(tSrBias.layout()));

FLASH_NAMESPACE::apply_mask</*Causal_mask=*/Is_causal, /*Has_mask=*/Has_mask, /*Has_bias=*/Has_bias>(
scores, /*mask=*/nullptr, bias, params.scale_softmax,
FLASH_NAMESPACE::apply_mask<Is_causal, Has_mask>(
scores, mask,
n_block * kBlockN + (tidx / 32 / AtomLayoutMS) * MMA_N_SdP * 16,
binfo.actual_seqlen_k,
m_block * kBlockM + get<0>(taccScS_row(0)),
binfo.actual_seqlen_q,
AtomLayoutMS * 16
);
} else {
FLASH_NAMESPACE::apply_mask</*Causal_mask=*/Is_causal, /*Has_mask=*/Has_mask, /*Has_bias=*/Has_bias>(
scores, /*mask=*/nullptr, /*bias=*/nullptr, params.scale_softmax,
FLASH_NAMESPACE::apply_mask<Is_causal, Has_mask>(
scores, /*mask=*/nullptr,
n_block * kBlockN + (tidx / 32 / AtomLayoutMS) * MMA_N_SdP * 16,
binfo.actual_seqlen_k,
m_block * kBlockM + get<0>(taccScS_row(0)),
Expand Down Expand Up @@ -965,8 +940,8 @@ inline __device__ void compute_dq_dk_dv_1colblock(const Params &params, const in
for (int i = 0; i < size(acc_dq); ++i) { atomicAdd(&tdQgdQaccum(i), acc_dq(i)); }
}
} else {
#pragma unroll
for (int i = 0; i < size(acc_dq); ++i) { acc_dq(i) *= params.scale_softmax; }
// #pragma unroll
// for (int i = 0; i < size(acc_dq); ++i) { acc_dq(i) *= params.scale_softmax; }
// Convert acc_dq from fp32 to fp16
Tensor rdQ = FLASH_NAMESPACE::convert_type<Element>(acc_dq);
Tensor taccdQrdQ = smem_thr_copy_dQ.retile_S(rdQ); // ((Atom, AtomNum), MMA_M, MMA_K)
Expand Down
4 changes: 2 additions & 2 deletions csrc/flash_dmattn/src/flash_bwd_preprocess_kernel.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -279,8 +279,8 @@ inline __device__ void convert_dQ(
for (int i = 0; i < size(acc_dq); ++i) { acc_dq(i) += tdQrdQaccum(i); }
tdQgdQaccum.data() = tdQgdQaccum.data() + params.dq_accum_split_stride;
}
#pragma unroll
for (int i = 0; i < size(acc_dq); ++i) { acc_dq(i) *= params.scale_softmax; }
// #pragma unroll
// for (int i = 0; i < size(acc_dq); ++i) { acc_dq(i) *= params.scale_softmax; }
// Convert acc_dq from fp32 to fp16
Tensor rdQ = FLASH_NAMESPACE::convert_type<Element>(acc_dq);
Tensor taccdQrdQ = smem_thr_copy_dQ.retile_S(rdQ); // ((Atom, AtomNum), MMA_N, MMA_N)
Expand Down
Loading