[RL] Release weight-transfer communicators without stopping inference (#61060)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Signed-off-by: fh3 <a@b>
This commit is contained in:
@@ -186,6 +186,7 @@ class SparseNCCLWeightTransferEngine(
|
||||
|
||||
def shutdown(self) -> None:
|
||||
if self.model_update_group is not None:
|
||||
self.model_update_group.destroy()
|
||||
self.model_update_group = None
|
||||
|
||||
|
||||
@@ -371,4 +372,6 @@ class SparseNCCLTrainerWeightTransferEngine(
|
||||
torch.cuda.current_stream().synchronize()
|
||||
|
||||
def shutdown(self) -> None:
|
||||
self.model_update_group = None
|
||||
if self.model_update_group is not None:
|
||||
self.model_update_group.destroy()
|
||||
self.model_update_group = None
|
||||
|
||||
Reference in New Issue
Block a user