Skip to content

cuda: add GGML_CUDA_BLOCKING_SYNC to eliminate 100% CPU busy-wait on full GPU offload #26487

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
hksk wants to merge 1 commit into
base: master
Choose a base branch
from
+16 −0
Open

cuda: add GGML_CUDA_BLOCKING_SYNC to eliminate 100% CPU busy-wait on full GPU offload #26487

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions ggml/src/ggml-cuda/ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -217,6 +217,22 @@ static int ggml_cuda_parse_id(char devName[]) {
static ggml_cuda_device_info ggml_cuda_init() {
ggml_cuda_device_info info = {};

// Optionally use blocking synchronization so that the host thread sleeps
// instead of busy-waiting while the GPU is working. This greatly reduces
// CPU usage for fully offloaded inference at the cost of a small increase
// in synchronization latency. Must be set before any CUDA context is
// created on the device.
const char * blocking_sync_env = getenv("GGML_CUDA_BLOCKING_SYNC");
if (blocking_sync_env != nullptr && std::atoi(blocking_sync_env) != 0) {
cudaError_t flags_err = cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync);
if (flags_err != cudaSuccess) {
GGML_LOG_WARN("%s: failed to set cudaDeviceScheduleBlockingSync: %s\n",
__func__, cudaGetErrorString(flags_err));
} else {
GGML_LOG_INFO("%s: using cudaDeviceScheduleBlockingSync\n", __func__);
}
}

cudaError_t err = cudaGetDeviceCount(&info.physical_device_count);
if (err != cudaSuccess) {
GGML_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
Expand Down