Skip to content

[TE] Restore original CP group after dynamic CP forward in TEDotProductAttention #5215

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 2 commits into from
Jun 11, 2026
Merged

[TE] Restore original CP group after dynamic CP forward in TEDotProductAttention #5215

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
bbb7deb
NCCL deadlock in dynamic-cp grpo training — TE context-parallel-group…
Jun 8, 2026
8be5af3
Fix lint errors (autoformat)
Jun 9, 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
17 changes: 17 additions & 0 deletions megatron/core/extensions/transformer_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1557,6 +1557,10 @@ def forward(
num_splits: Optional[int] = None,
) -> torch.Tensor:
"""Forward."""
# Save TE's current CP group before potential DCP switch (for restore at end).
_te_orig_cp_group = self.cp_group
_te_orig_cp_global_ranks = self.cp_global_ranks

if packed_seq_params is not None:
# If Dynamic CP group is provided, update TE DPA CP group
if packed_seq_params.local_cp_size is not None:
Expand Down Expand Up @@ -1647,6 +1651,19 @@ def forward(
_fa_kwargs["num_splits"] = num_splits
core_attn_out = super().forward(query, key, value, attention_mask, **_fa_kwargs)

# Restore TE's CP group after dynamic CP forward.
if (
packed_seq_params is not None
and packed_seq_params.local_cp_size is not None
and self.config.context_parallel_size > 1
):
super().set_context_parallel_group(
_te_orig_cp_group,
_te_orig_cp_global_ranks,
TEDotProductAttention.cp_stream,
self.cp_comm_type,
)

return core_attn_out

def sharded_state_dict(
Expand Down
Loading