CUDA: fix FA VKQ accumulator overflow (#17746)

This commit is contained in:
Johannes Gäßler
2025-12-05 09:18:10 +01:00
committed by GitHub
parent 668ed76574
commit e95d0bc8fd
5 changed files with 11 additions and 5 deletions
+1 -1
View File
@@ -572,7 +572,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) ?
slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] + FATTN_KQ_MAX_OFFSET);
}
}