Skip to content

Fix NCCL Error 7 by removing eager_connect_single_device #320

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
d4l3k wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix NCCL Error 7 by removing eager_connect_single_device #320

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
431f4ef
Fix NCCL non-blocking test failures by adding cuda.synchronize()
d4l3k Mar 24, 2026
d2888ad
Fix NCCL Error 7 by removing eager_connect_single_device
d4l3k Mar 24, 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
10 changes: 7 additions & 3 deletions torchft/process_group.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -864,9 +864,13 @@ def _create_pg(self, store: Store, rank: int, world_size: int) -> BaseProcessGro
# pyre-fixme[16]: no attribute ProcessGroupNCCL
backend_class = BaseProcessGroupNCCL(store, rank, world_size, opts)
backend_class._set_sequence_number_for_group()
backend_class.eager_connect_single_device(
torch.device(torch.accelerator.current_device_index())
)
# NOTE: We intentionally do NOT call eager_connect_single_device here.
# In non-blocking mode (blocking=False), eager_connect starts an async
# communicator init that cannot be waited on through Python APIs. The
# first collective would then fail with "NCCL Error 7: operation in
# progress". Instead, we let the communicator be lazily initialized by
# the first collective, which properly handles the async init inside
# its ncclGroupStart/ncclGroupEnd context.
pg._register_backend(
torch.device("cuda"), ProcessGroup.BackendType.NCCL, backend_class
)
Expand Down
Loading