tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-14 04:17:53 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: use mma FA kernel for gqa > 4 on RTX 4000 (#15035)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-08-02 16:37:08 +02:00
committed by GitHub
parent 3303c19b16
commit 03d4698218
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
ggml/src/ggml-cuda/fattn.cu
View File

@@ -315,8 +315,9 @@ 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_bs1 = new_mma_available(cc) && gqa_opt_applies &&
(Q->ne[3] > 1 || cc < GGML_CUDA_CC_ADA_LOVELACE) && !mma_needs_data_conversion;
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 &&
(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) {
if (prec == GGML_PREC_DEFAULT) {