EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

CUDA: app option to compile without FlashAttention (#12025)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-02-22 20:44:34 +01:00
committed by GitHub
parent 36c258ee92
commit a28e0d5eb1
13 changed files with 46 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-cuda/fattn-vec-f32.cuh
View File

@@ -41,10 +41,7 @@ static __global__ void flash_attn_vec_ext_f32(
const int ne1,
const int ne2,
const int ne3) {
#ifndef FLASH_ATTN_AVAILABLE
NO_DEVICE_CODE;
return;
#endif // FLASH_ATTN_AVAILABLE
#ifdef FLASH_ATTN_AVAILABLE
// Skip unused kernel variants for faster compilation:
if (use_logit_softcap && !(D == 128 || D == 256)) {
@@ -281,6 +278,9 @@ static __global__ void flash_attn_vec_ext_f32(
if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
}
#else
NO_DEVICE_CODE;
#endif // FLASH_ATTN_AVAILABLE
}
template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>