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

CUDA: fix overflow in FA, tune performance (#14840)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-07-23 21:43:25 +02:00
committed by GitHub
parent b284197df4
commit a86f52b285
8 changed files with 98 additions and 246 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn.cu
+3 -13
View File
@@ -280,22 +280,12 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
const enum ggml_prec prec = ggml_flash_attn_ext_get_prec(KQV);
if (GGML_CUDA_CC_IS_AMD(cc)) {
#if defined(GGML_HIP_ROCWMMA_FATTN)
if (fp16_mma_available(cc)) {
ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
return;
}
#endif // defined(GGML_HIP_ROCWMMA_FATTN)
// On AMD the tile kernels perform poorly, use the vec kernel instead:
if (prec == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
} else {
ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
}
if (GGML_CUDA_CC_IS_AMD(cc) && fp16_mma_available(cc)) {
ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
return;
}
#endif // defined(GGML_HIP_ROCWMMA_FATTN)
if (!fast_fp16_available(cc)) {
if (Q->ne[1] <= 8 || Q->ne[0] == 256) {