Skip to content

[release/2.8] Using c10d.barrier() in test_extra_cuda_context test in test_c10d_nccl.py #2522

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

Merged
merged 1 commit into from
Aug 15, 2025
Merged

[release/2.8] Using c10d.barrier() in test_extra_cuda_context test in test_c10d_nccl.py #2522

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
8 changes: 8 additions & 0 deletions test/distributed/test_c10d_nccl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -639,6 +639,14 @@ def _helper_test_extra_cuda_context_by_memory(self):
"""
device = torch.device(f"cuda:{self.rank:d}")
x = torch.empty((1,), device=device)

# We need this barrier to ensure that all nodes have completed init_process_group
# If rank=0 gets a mem snapshot before other nodes have finished init_process_group,
# then we artificially see a bump in memory usage. As per the following comment,
# we are going to be moving away from this function:
# https://github.com/pytorch/pytorch/pull/154174#discussion_r2105065931
c10d.barrier()

# Rank 0 takes a snapshot before collective -- this snapshot should have
# included rank 0's own context.
if self.rank == 0:
Expand Down