mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-11-20 12:07:33 +00:00
CANN: Optimize ggml_cann_set_device (#15935)
* CANN: Fix ggml_cann_set_device to avoid redundant device switches - Added a check to skip aclrtSetDevice if the current device is already set. - Prevents unnecessary context switches while keeping thread/device consistency. * CANN: add device default id
This commit is contained in:
@@ -526,7 +526,10 @@ struct ggml_backend_cann_context {
|
||||
*/
|
||||
aclrtStream stream(int stream) {
|
||||
if (streams[stream] == nullptr) {
|
||||
ggml_cann_set_device(device);
|
||||
// If the device is not set here, destroying the stream later may cause a mismatch
|
||||
// between the thread contexts where the stream was created and destroyed.
|
||||
// However, I printed the device_id, thread_id, and stream, and they are all consistent.
|
||||
ACL_CHECK(aclrtSetDevice(device));
|
||||
ACL_CHECK(aclrtCreateStream(&streams[stream]));
|
||||
}
|
||||
return streams[stream];
|
||||
|
||||
Reference in New Issue
Block a user