CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-10 10:27:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: avoid mul + bias fusion when doing fusion (#16935)

Browse Source
This commit is contained in:
Aman Gupta
2025-11-04 10:53:48 +08:00
committed by GitHub
parent c5023daf60
commit 2759ccdb4a
1 changed files with 17 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -2115,6 +2115,14 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_f(const ggml_tensor * tensor) {
const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
use_mul_mat_vec_f = use_mul_mat_vec_f && ggml_cuda_should_use_mmvf(src0->type, cc, src0->ne, is_mul_mat_id ? src1->ne[2] : src1->ne[1]);
const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
ggml_backend_buft_is_cuda_split(src1->buffer->buft);
//TODO: add support for fusion for split buffers
if (split) {
return false;
}
//we only support fusion for ncols_dst = 1
if (tensor->op == GGML_OP_MUL_MAT && dst->ne[1] != 1) {
return false;
@@ -2154,6 +2162,15 @@ static bool ggml_cuda_should_fuse_mul_mat_vec_q(const ggml_tensor * tensor) {
return false;
}
const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft) ||
ggml_backend_buft_is_cuda_split(src1->buffer->buft);
//TODO: add support for fusion for split buffers
if (split) {
return false;
}
return use_mul_mat_vec_q;
}