vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

cuda : fix disabling device with --tensor-split 1,0

Browse source
This commit is contained in:
Jared Van Bortel 2023-11-05 00:56:32 -04:00
parent f28af0d81a
commit 05c51f96fe
1 changed files with 3 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

3
ggml-cuda.cu
View file

@ -7117,6 +7117,9 @@ static void ggml_cuda_op_mul_mat(
CUDA_CHECK(ggml_cuda_set_device(g_main_device));
for (int64_t id = 0; id < g_device_count; ++id) {
if (g_tensor_split[id] >= (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
continue;
}
for (int64_t is = 0; is < is_max; ++is) {
CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
}