ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-02-16 07:10:22 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

CUDA: fix Volta FlashAttention logic (llama/11615)

Browse Source
This commit is contained in:
Johannes Gäßler 2025-02-03 13:25:56 +01:00 committed by Georgi Gerganov
parent fad2806352
commit dbeb7916b8
2 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-cuda/fattn-wmma-f16.cu
View File

@ -561,7 +561,7 @@ void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_ten
ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst); ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
break; break;
// case 256: // case 256:
// ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst); // ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, float>(ctx, dst);
// break; // break;
default: default:
GGML_ABORT("fatal error"); GGML_ABORT("fatal error");

3
ggml/src/ggml-cuda/fattn.cu
View File

@ -235,7 +235,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
return; return;
} }
if (!new_mma_available(cc)) { if (!fp16_mma_available(cc)) {
if (prec == GGML_PREC_DEFAULT) { if (prec == GGML_PREC_DEFAULT) {
if (Q->ne[1] <= 8) { if (Q->ne[1] <= 8) {
ggml_cuda_flash_attn_ext_vec_f16(ctx, dst); ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
@ -265,6 +265,7 @@ 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: // The MMA implementation needs Turing or newer, use the old WMMA code for Volta:
if (cc == GGML_CUDA_CC_VOLTA) { if (cc == GGML_CUDA_CC_VOLTA) {
ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst); ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
return;
} }
ggml_cuda_flash_attn_ext_mma_f16(ctx, dst); ggml_cuda_flash_attn_ext_mma_f16(ctx, dst);