Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16 (#15131)

Browse Source 
* CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16
This commit is contained in:
Johannes Gäßler
2025-08-07 10:53:21 +02:00
committed by GitHub
parent 20638e4f16
commit 1d72c84188
13 changed files with 750 additions and 225 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn.cu
+2 -2
View File
@@ -327,7 +327,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
const bool gqa_opt_applies = ((Q->ne[2] / K->ne[2]) % 2 == 0) && mask; // The mma-based kernels have GQA-specific optimizations
const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
const bool mma_faster_for_rtx4000 = Q->ne[3] > 1 || (Q->ne[2] > 4*K->ne[2] && K->ne[1] >= 8192);
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
const bool mma_faster_for_bs1 = turing_mma_available(cc) && gqa_opt_applies && !mma_needs_data_conversion &&
(cc < GGML_CUDA_CC_ADA_LOVELACE || mma_faster_for_rtx4000);
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*warp_size) == 0;
if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1) {
@@ -340,7 +340,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
}
// The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
if (fp16_mma_available(cc) && !new_mma_available(cc)) {
if (fp16_mma_available(cc) && !turing_mma_available(cc)) {
ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
return;
}