Skip to content

[Dev] Remove redundant reduce in aux_loss logging #2094

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
BestJuly merged 4 commits into from
Nov 12, 2025
Merged

[Dev] Remove redundant reduce in aux_loss logging #2094

Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0d0ddf3
Remove redundant reduce in aux_loss logging
BestJuly Nov 3, 2025
2c6ec4c
Merge branch 'dev' into lit/remove_duplicated_aux_loss_reduce_dev
BestJuly Nov 4, 2025
d0d5c49
Fix comment
BestJuly Nov 4, 2025
6f26d81
Merge branch 'dev' into lit/remove_duplicated_aux_loss_reduce_dev
BestJuly Nov 4, 2025
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
15 changes: 9 additions & 6 deletions megatron/core/transformer/moe/moe_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -770,12 +770,15 @@ def reduce_aux_losses_tracker_across_ranks(track_names: Optional[List[str]] = No
torch.distributed.all_reduce(
values, group=tracker[name]['avg_group'], op=torch.distributed.ReduceOp.AVG
)
# This ensures proper loss averaging across all ranks including CP ranks
torch.distributed.all_reduce(
values,
group=parallel_state.get_data_parallel_group(with_context_parallel=True),
op=torch.distributed.ReduceOp.AVG,
)
# Average aux losses across data parallel ranks.
# The `global_load_balancing_loss` already uses `tp_dp_cp_group` in `reduce_group`,
# so we don't need to reduce it again. Others use tp_cp_group in `reduce_group`.
if name != "global_load_balancing_loss":
torch.distributed.all_reduce(
values,
group=parallel_state.get_data_parallel_group(with_context_parallel=False),
op=torch.distributed.ReduceOp.AVG,
)


def track_moe_metrics(
Expand Down
Loading