tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-26 03:03:25 -04:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: fix FA out-of-bounds writes (#7465)

Browse Source
This commit is contained in:
Johannes Gäßler
2024-05-22 17:58:25 +02:00
committed by GitHub
parent b18532a4ef
commit 38c03478a3
4 changed files with 18 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml-cuda/fattn-tile-f16.cu
View File

@ -238,6 +238,10 @@ static __global__ void flash_attn_tile_ext_f16(
for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
const int j_VKQ = j_VKQ_0 + threadIdx.y;
if (ic0 + j_VKQ >= ne01) {
return;
}
half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
kqsum_j = warp_reduce_sum(kqsum_j);