Skip to content

fix ddp bug when --overlap-grad-reduce and --num-optim > 1 #3693

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
wplf wants to merge 3 commits into
base: main
Choose a base branch
from
+3 −3
Open

fix ddp bug when --overlap-grad-reduce and --num-optim > 1 #3693

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
d8d94ab
fix ddp bug when --overlap-grad-reduce and --num-distributed-optimize…
wplf Mar 4, 2026
370e4a2
Merge branch 'main' into fix-ddp-bug-with-overlap
wplf Mar 4, 2026
7220436
Merge branch 'main' into fix-ddp-bug-with-overlap
wplf Mar 5, 2026
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
6 changes: 3 additions & 3 deletions megatron/core/distributed/param_and_grad_buffer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -419,10 +419,10 @@ def start_grad_sync(self, force_all_reduce: Optional[bool] = False):
# need to overlap communication.
stream_context = torch.cuda.stream(self.communication_stream)

# The RS/AR communication stream needs to wait for the default stream
# The RS/AR communication stream needs to wait for the current stream
# to complete its gradient computation before launching the next
# gradient reduction collective.
self.communication_stream.wait_stream(torch.cuda.default_stream())
self.communication_stream.wait_stream(torch.cuda.current_stream())
else:
stream_context = nullcontext()

Expand Down Expand Up @@ -529,7 +529,7 @@ def finish_grad_sync(self, force_all_reduce: Optional[bool] = False):
# When using multiple DistOpt instances, we don't need to sync here as we launch
# communications on a separate communication stream.
if self.ddp_config.num_distributed_optimizer_instances > 1:
torch.cuda.default_stream().wait_stream(self.communication_stream)
torch.cuda.current_stream().wait_stream(self.communication_stream)
return
assert self.grad_reduce_handle is not None, (
f"Communication call has not been issued for this bucket "
Expand Down