From 1aa87ee53d05505247c54612e40f6a38c680b433 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E8=95=AD=E6=BE=A7=E9=82=A6?= <45505768+shou692199@users.noreply.github.com> Date: Fri, 21 Mar 2025 14:58:47 +0800 Subject: [PATCH 1/6] [SYCL] Fix build on Windows when ccache enabled (#9954) (#9976) * [SYCL] Fix build on Windows when ccache enabled (#9954) * take effect only on windows and force it to icl --------- Co-authored-by: Romain Biessy --- ggml/src/CMakeLists.txt | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index c1c7498694beb..1e4c2422756ac 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -76,7 +76,11 @@ if (GGML_CCACHE) set(GGML_CCACHE_VARIANT sccache) endif() # TODO: should not be set globally - set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}") + if (GGML_SYCL AND GGML_CCACHE_FOUND AND WIN32) + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "ccache compiler_type=icl") + else () + set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE "${GGML_CCACHE_VARIANT}") + endif () set(ENV{CCACHE_SLOPPINESS} time_macros) message(STATUS "${GGML_CCACHE_VARIANT} found, compilation results will be cached. Disable with GGML_CCACHE=OFF.") else() From ea1518e839abe668c20f7e0074c0721f803da898 Mon Sep 17 00:00:00 2001 From: marcoStocchi Date: Fri, 21 Mar 2025 10:12:45 +0100 Subject: [PATCH 2/6] llama-tts : avoid crashes related to bad model file paths (#12482) --- examples/tts/tts.cpp | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/examples/tts/tts.cpp b/examples/tts/tts.cpp index d953cadd62dcf..4cc42e1674ccc 100644 --- a/examples/tts/tts.cpp +++ b/examples/tts/tts.cpp @@ -571,6 +571,10 @@ int main(int argc, char ** argv) { model_ttc = llama_init_ttc.model.get(); ctx_ttc = llama_init_ttc.context.get(); + if (model_ttc == nullptr || ctx_ttc == nullptr) { + return ENOENT; + } + const llama_vocab * vocab = llama_model_get_vocab(model_ttc); // TODO: refactor in a common struct @@ -586,6 +590,10 @@ int main(int argc, char ** argv) { model_cts = llama_init_cts.model.get(); ctx_cts = llama_init_cts.context.get(); + if (model_cts == nullptr || ctx_cts == nullptr) { + return ENOENT; + } + std::vector smpl(n_parallel); for (int i = 0; i < n_parallel; ++i) { params.sampling.no_perf = (i != 0); From 960e72607761eb2dd170b33f02a5a2840ec412fe Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= Date: Fri, 21 Mar 2025 10:21:36 +0100 Subject: [PATCH 3/6] chore : cleanup llama_model_loader::TENSOR_ usage (#12492) --- src/llama-model.cpp | 42 +++++++++++++++++++++--------------------- 1 file changed, 21 insertions(+), 21 deletions(-) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index cd7e0a0c4dbf8..9ccfc7fc61c47 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -2329,7 +2329,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0); layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), { n_embd }, 0); - layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, TENSOR_NOT_REQUIRED); if (layer.wqkv == nullptr) { layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0); layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0); @@ -3215,16 +3215,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) { auto & layer = layers[i]; layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); - layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED); if (layer.wqkv == nullptr) { layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); - layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, TENSOR_NOT_REQUIRED); } layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); @@ -3335,12 +3335,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {time_mix_extra_dim, n_embd, 5}, 0); layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0); - layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_w = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_W, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_K, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_v = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_V, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_r = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_R, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_g = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_G, "weight", i), {n_embd, 1, 1}, TENSOR_NOT_REQUIRED); + layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, TENSOR_NOT_REQUIRED); GGML_ASSERT(!(layer.time_mix_lerp_fused == NULL && layer.time_mix_lerp_w == NULL)); layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, 0); @@ -3370,7 +3370,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); - output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, TENSOR_NOT_REQUIRED); output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); const int time_mix_extra_dim = hparams.time_mix_extra_dim; @@ -3396,7 +3396,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_lerp_x = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_X, "weight", i), {n_embd, 1, 1}, 0); layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0); - layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_first = create_tensor(tn(LLM_TENSOR_TIME_MIX_FIRST, "weight", i), {head_size, n_embd / head_size}, TENSOR_NOT_REQUIRED); layer.time_mix_decay = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY, "weight", i), {n_embd}, 0); layer.time_mix_decay_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W1, "weight", i), {n_embd, time_decay_extra_dim}, 0); layer.time_mix_decay_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_DECAY_W2, "weight", i), {time_decay_extra_dim, attn_hidden_size}, 0); @@ -3405,9 +3405,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); layer.time_mix_gate = create_tensor(tn(LLM_TENSOR_TIME_MIX_GATE, "weight", i), {attn_hidden_size, n_embd}, 0); // optional bias tensors - layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_key_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED); + layer.time_mix_value_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "bias", i), {attn_key_value_size}, TENSOR_NOT_REQUIRED); + layer.time_mix_receptance_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "bias", i), {attn_hidden_size}, TENSOR_NOT_REQUIRED); layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); @@ -3528,8 +3528,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0); } - layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED); + layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED); try { layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0); @@ -3546,8 +3546,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0); layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0); - layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); - layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED); + layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED); + layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED); layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); From af04481e6b3b8dcdf5c1cc7d84bc7ece5658e9ab Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 21 Mar 2025 16:14:29 +0200 Subject: [PATCH 4/6] model : do not repack if a GPU device is present (#12498) ggml-ci --- src/llama-model.cpp | 33 +++++++++++++++++++++++---------- 1 file changed, 23 insertions(+), 10 deletions(-) diff --git a/src/llama-model.cpp b/src/llama-model.cpp index 9ccfc7fc61c47..26ac5e99bfc7a 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -271,19 +271,32 @@ static buft_list_t make_cpu_buft_list(const std::vector & de } } - // add extra buffer types - auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); - auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); - auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) - ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); - if (ggml_backend_dev_get_extra_bufts_fn) { - ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); - while (extra_bufts && *extra_bufts) { - buft_list.emplace_back(cpu_dev, *extra_bufts); - ++extra_bufts; + bool has_gpu_device = false; + for (auto * dev : devices) { + if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) { + has_gpu_device = true; + break; } } + // add extra buffer types, only if no GPU device is present + // ref: https://github.com/ggml-org/llama.cpp/issues/12481#issuecomment-2743136094 + if (!has_gpu_device) { + auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU); + auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev); + auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) + ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts"); + if (ggml_backend_dev_get_extra_bufts_fn) { + ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev); + while (extra_bufts && *extra_bufts) { + buft_list.emplace_back(cpu_dev, *extra_bufts); + ++extra_bufts; + } + } + } else { + LLAMA_LOG_WARN("%s: disabling extra buffer types (i.e. repacking) since a GPU device is available\n", __func__); + } + // add a host buffer type // storing the tensors in a host buffer is useful when the processing of large batches // is offloaded to a GPU device, since it reduces the time spent on data transfers From 30c42ef5cbb2db756eff9aadc6b6c528ba68388d Mon Sep 17 00:00:00 2001 From: Eve <139727413+netrunnereve@users.noreply.github.com> Date: Fri, 21 Mar 2025 19:27:47 +0000 Subject: [PATCH 5/6] vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472) --- .../ggml-vulkan/vulkan-shaders/dequant_funcs.comp | 4 ++-- .../vulkan-shaders/mul_mat_vec_iq2_s.comp | 4 ++-- .../vulkan-shaders/mul_mat_vec_iq3_s.comp | 2 +- ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp | 14 +++++++------- 4 files changed, 12 insertions(+), 12 deletions(-) diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp index 8835c442ecfd8..2a162a2c81543 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.comp @@ -82,8 +82,8 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) { return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])); } vec4 dequantize4(uint ib, uint iqs, uint a_offset) { - const i8vec2 v0 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2]); - const i8vec2 v1 = unpack8(data_a_packed16[a_offset + ib].qs[iqs/2 + 1]); + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[a_offset + ib].qs[iqs/2 + 1])).xy; return vec4(v0.x, v0.y, v1.x, v1.y); } #endif diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp index 9718a05e5adb2..8d01536fa69c0 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp @@ -19,8 +19,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const float db = d * (0.5 + scale) * 0.25; const uint qh = data_a[ibi].qh[ib32]; - const u8vec2 qs16 = unpack8(data_a_packed16[ibi].qs[itid]); - const u8vec2 sign16 = unpack8(data_a_packed16[ibi].qs[QUANT_K / 16 + itid]); + const u8vec2 qs16 = unpack8(uint32_t(data_a_packed16[ibi].qs[itid])).xy; // vec4 used due to #12147 + const u8vec2 sign16 = unpack8(uint32_t(data_a_packed16[ibi].qs[QUANT_K / 16 + itid])).xy; [[unroll]] for (uint l = 0; l < 2; ++l) { const uint8_t sign = sign16[l]; const uint qs = qs16[l] | ((qh << (8 - nibble_shift - 2 * l)) & 0x300); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp index af48f32902fe2..f021e40476199 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp @@ -21,7 +21,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, sum[j] = 0.0; } [[unroll]] for (uint l = 0; l < 4; ++l) { - const u8vec2 qs = unpack8(data_a_packed16[ibi].qs[4 * ib32 + l]); + const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147 const uint sign = data_a[ibi].signs[4 * ib32 + l]; const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)])); const vec4 grid1 = vec4(unpack8(iq3s_grid[qs.y | ((qh << (7 - 2*l)) & 0x100)])); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp index 0d03411f24ca4..5a0054bac336c 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp @@ -336,8 +336,8 @@ void main() { const uint iqs = idx & 0x07; const float d = float(data_a_packed16[ib].d); - const i8vec2 v0 = unpack8(data_a_packed16[ib].qs[2*iqs]); - const i8vec2 v1 = unpack8(data_a_packed16[ib].qs[2*iqs + 1]); + const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147 + const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy; const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d; buf_a[buf_idx ] = FLOAT_TYPE(v.x); @@ -544,7 +544,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -564,7 +564,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -586,7 +586,7 @@ void main() { const float db = d * 0.25 * (0.5 + scale); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1]; - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid)); + const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -611,7 +611,7 @@ void main() { const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4)); const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign)))); const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); @@ -631,7 +631,7 @@ void main() { const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign))); const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf)); const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2)); - const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); + const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147 buf_a[buf_idx ] = FLOAT_TYPE(v.x); buf_a[buf_idx + 1] = FLOAT_TYPE(v.y); From 4375415b4abf94fb36a5fd15f233ac0ee23c0bd1 Mon Sep 17 00:00:00 2001 From: stduhpf Date: Fri, 21 Mar 2025 20:34:50 +0100 Subject: [PATCH 6/6] Vulkan: RTE rounding for cpy to quant (#12480) * Vulkan: RTE rounding for cpy to quant Co-Authored-By: Jeff Bolz * remove trailing whitespace * avoid duplicating pipeline_cpy_f32_quant * fix copypasting issue * remove duplicated code --------- Co-authored-by: Jeff Bolz --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 22 +++++++++++++------ .../vulkan-shaders/copy_to_quant.comp | 5 +++++ .../vulkan-shaders/vulkan-shaders-gen.cpp | 1 + 3 files changed, 21 insertions(+), 7 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index d450fe9a2f2f6..649504566ab58 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -2281,13 +2281,21 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f32, "contig_cpy_f32_f32", contig_cpy_f32_f32_len, contig_cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f32_f16, "contig_cpy_f32_f16", contig_cpy_f32_f16_len, contig_cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_contig_cpy_f16_f16, "contig_cpy_f16_f16", contig_cpy_f16_f16_len, contig_cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1); - - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); - ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + if (device->float_controls_rte_fp16) { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_rte_len, cpy_f32_q4_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_rte_len, cpy_f32_q4_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_rte_len, cpy_f32_q5_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_rte_len, cpy_f32_q5_1_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_rte_len, cpy_f32_q8_0_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_rte_len, cpy_f32_iq4_nl_rte_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } else { + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_0], "cpy_f32_q4_0", cpy_f32_q4_0_len, cpy_f32_q4_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q4_1], "cpy_f32_q4_1", cpy_f32_q4_1_len, cpy_f32_q4_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_0], "cpy_f32_q5_0", cpy_f32_q5_0_len, cpy_f32_q5_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q5_1], "cpy_f32_q5_1", cpy_f32_q5_1_len, cpy_f32_q5_1_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q5_1), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_Q8_0], "cpy_f32_q8_0", cpy_f32_q8_0_len, cpy_f32_q8_0_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q8_0), 1, 1}, {}, 1); + ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_quant[GGML_TYPE_IQ4_NL], "cpy_f32_iq4_nl", cpy_f32_iq4_nl_len, cpy_f32_iq4_nl_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_IQ4_NL), 1, 1}, {}, 1); + } ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_0], "cpy_q4_0_f32", cpy_q4_0_f32_len, cpy_q4_0_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_0), 1, 1}, {}, 1); ggml_vk_create_pipeline(device, device->pipeline_cpy_quant_f32[GGML_TYPE_Q4_1], "cpy_q4_1_f32", cpy_q4_1_f32_len, cpy_q4_1_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {(uint32_t)ggml_blck_size(GGML_TYPE_Q4_1), 1, 1}, {}, 1); diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp index c813f14044eca..9c76437d9b0b9 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp @@ -1,5 +1,10 @@ #version 450 +#if RTE16 +#extension GL_EXT_spirv_intrinsics : enable +spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits +#endif // RTE16 + #include "types.comp" #include "generic_unary_head.comp" diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp index eb2ad63ff6bf0..519e610e31dc6 100644 --- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp @@ -445,6 +445,7 @@ void process_shaders() { for (std::string t : {"q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) { string_to_spv("cpy_f32_" + t, "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); + string_to_spv("cpy_f32_" + t + "_rte", "copy_to_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}}); string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}}); }