CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: skip masked KV slices for all FA kernels (#14924)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-07-30 15:46:13 +02:00
committed by GitHub
parent 00131d6eaf
commit 92b8810ec7
9 changed files with 120 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-cuda/common.cuh
View File

@@ -432,6 +432,20 @@ static __global__ void reduce_rows_f32(const float * x, float * dst, const int n
dst[row] = norm ? sum / ncols : sum;
}
template<int width = WARP_SIZE>
static __device__ __forceinline__ int warp_reduce_all(int x) {
#ifdef GGML_USE_HIP
#pragma unroll
for (int offset = width/2; offset > 0; offset >>= 1) {
x = x && __shfl_xor_sync(0xffffffff, x, offset, width);
}
return x;
#else
static_assert(width == WARP_SIZE, "width != WARP_SIZE not implemented");
return __all_sync(0xffffffff, x);
#endif // GGML_USE_HIP
}
template<int width = WARP_SIZE>
static __device__ __forceinline__ float warp_reduce_max(float x) {
#pragma unroll