upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-08-06 18:10:51 -05:00
Code Issues Packages Projects Releases Wiki Activity

allow disabling CUDA reductions completely (falling back to the non-CUDA butterfly mode)

Browse Source
This commit is contained in:
Scott Cutler
2026-04-24 12:58:19 -07:00
parent 7449c1d30a
commit 2b91e21d14
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -1218,6 +1218,16 @@ static void * ggml_backend_cuda_comm_init(ggml_backend_t * backends, size_t n_ba
}
}
// GGML_CUDA_ALLREDUCE=none disables the CUDA-specific AllReduce entirely,
// so the meta-backend falls back to its generic butterfly reduction.
{
const char * env = getenv("GGML_CUDA_ALLREDUCE");
if (env != nullptr && strcmp(env, "none") == 0) {
GGML_LOG_INFO("%s: GGML_CUDA_ALLREDUCE=none; using meta-backend butterfly reduction\n", __func__);
return nullptr;
}
}
std::vector<int> dev_ids;
dev_ids.reserve(n_backends);
for (size_t i = 0; i < n_backends; i++) {