Skip to content

Fix: nccl-ub in ddp path #3181

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
yaox12 merged 3 commits into from
Feb 1, 2026
Merged

Fix: nccl-ub in ddp path #3181

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c7039b9
fix
youngeunkwon0405 Jan 30, 2026
8a62df3
Address review's comment
youngeunkwon0405 Jan 30, 2026
13376d1
Merge branch 'dev' into fix_nccl_ub_in_ddp
yaox12 Feb 1, 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: 6 additions & 0 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 @@ -753,6 +753,12 @@ def _does_param_require_new_bucket(param):
group=self.data_parallel_group,
symmetric=not self.ddp_config.disable_symmetric_registration,
)
# Since nccl communicator group is created lazily, we need to perform a warmup call to
# initialize NCCL comm buffers for this dp_group before doing buffer registration.
torch.distributed.barrier()
tmp_warmup_tensor = torch.zeros([1], device="cuda")
torch.distributed.all_reduce(tmp_warmup_tensor, group=self.data_parallel_group)
torch.distributed.barrier()
else:
# If nccl_ub is False, mem_alloc_context is nullcontext.
mem_alloc_context = nullcontext
Expand Down
Loading