Skip to content

[sync] Fix bug in param_norm computation where some ranks might call collective and some might not #852

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 3 commits into from
Oct 14, 2025
Merged

[sync] Fix bug in param_norm computation where some ranks might call collective and some might not #852

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
fecfaef
sync with mr 3918
ananthsub Oct 2, 2025
73acd67
add unit tests
ananthsub Oct 13, 2025
d8e2d1d
rebase
ananthsub Oct 13, 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
18 changes: 10 additions & 8 deletions src/megatron/bridge/training/utils/train_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,14 +186,16 @@ def calc_params_l2_norm(
False, # no per-parameter norm.
)
sharded_norm_2 = sharded_norm * sharded_norm
# Sum over all DP groups, including CP since distributed optimizer state is
# sharded jointly over DP+CP.
torch.distributed.all_reduce(
sharded_norm_2,
op=torch.distributed.ReduceOp.SUM,
group=parallel_state.get_data_parallel_group(with_context_parallel=True),
)
norm_2 += sharded_norm_2
else:
sharded_norm_2 = torch.zeros((1,), dtype=torch.float32, device="cuda")
# Sum over all DP groups, including CP since distributed optimizer state is
# sharded jointly over DP+CP.
torch.distributed.all_reduce(
sharded_norm_2,
op=torch.distributed.ReduceOp.SUM,
group=parallel_state.get_data_parallel_group(with_context_parallel=True),
)
norm_2 += sharded_norm_2

# Add norm contribution from expert layers in MoEs.
if len(moe_params_data) > 0:
Expand Down
Loading
Loading