CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-30 08:42:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: fix half2 -> half conversion for HIP (#15529)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-08-23 21:37:06 +02:00
committed by GitHub
parent 611f419cff
commit 710dfc465a
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

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

@@ -258,7 +258,7 @@ static __global__ void flash_attn_tile_ext_f16(
const half val = hexp(sink - kqmax[j0/nwarps]); const half val = hexp(sink - kqmax[j0/nwarps]);
kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale; kqsum[j0/nwarps] = kqsum[j0/nwarps] * KQ_max_scale;
if (threadIdx.x == 0) { if (threadIdx.x == 0) {
kqsum[j0/nwarps].x = __hadd(kqsum[j0/nwarps].x, val); kqsum[j0/nwarps].x = __hadd(__low2half(kqsum[j0/nwarps]), val);
} }
#pragma unroll #pragma unroll