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

CUDA: Add Flash Attention Support for Head Dimension 512 (#20998)

Browse Source 
* flash attention support for head dimension 512 added

* FA D=512 - match 576 configs, limit ncols2, revert vec cap

* fix HIP tile kernel build for D=512

* fix HIP tile kernel occupancy for D=512 on AMD

* Apply suggestions from code review

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>

* fix tile FA compilation

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Anav Prasad
2026-04-01 09:07:24 +02:00
committed by GitHub
co-authored by Johannes Gäßler
parent 4951250235
commit 88458164c7
14 changed files with 86 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn.cu
+8 -3
View File
@@ -135,6 +135,10 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
GGML_ASSERT(V->ne[0] == 256);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<256, 256>(ctx, dst);
break;
case 512:
GGML_ASSERT(V->ne[0] == 512);
ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2<512, 512>(ctx, dst);
break;
case 576: {
// For Deepseek, go straight to the ncols1 switch to avoid compiling unnecessary kernels.
GGML_ASSERT(V->ne[0] == 512);
@@ -336,7 +340,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
case 128:
case 112:
case 256:
if (V->ne[0] != K->ne[0]) {
case 512:
if (!gqa_opt_applies) {
return BEST_FATTN_KERNEL_NONE;
}
break;
@@ -424,7 +429,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
}
// Use the WMMA kernel if possible:
if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 576) {
if (ggml_cuda_should_use_wmma_fattn(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 512 && Q->ne[0] != 576) {
if (can_use_vector_kernel && Q->ne[1] <= 2) {
return BEST_FATTN_KERNEL_VEC;
}
@@ -457,7 +462,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
}
// Use MFMA flash attention for CDNA (MI100+):
if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 576) {
if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72 && Q->ne[0] != 256 && Q->ne[0] != 512 && Q->ne[0] != 576) {
const int64_t eff_nq = Q->ne[1] * (gqa_opt_applies ? gqa_ratio : 1);
// MMA vs tile crossover benchmarked on MI300X @ d32768:
// hsk=64 (gqa=4): MMA wins at eff >= 128 (+11%)