test(clip_grads): add regression coverage for get_grad_norm_fp32

[Mattral]

Summary

Follow-up to #5530. test_clip_grads.py had no test coverage for
get_grad_norm_fp32 itself — this adds regression tests for all three
norm_type branches, including the empty-grads_for_norm case that #5530 fixed.

Changes

tests/unit_tests/optimizer/test_clip_grads.py

• test_empty_grads_returns_zero (parametrized over 2.0, inf, 1.0):
  regression test for the crash fixed in fix(clip_grads): handle empty grads_for_norm in inf-norm and p-norm paths #5530
• test_nonempty_grads_matches_reference (same parametrization): baseline
  correctness check against a plain torch.norm computation

Testing

• Verified locally via linting/formatting (black, isort, pylint, ruff —
  all clean via tools/autoformat.sh)
• Could not execute pytest locally (no GPU available in this dev environment,
  and every test in this suite requires torch.distributed initialization);
  relying on CI's GPU runners for actual execution
• No changes to megatron/core/optimizer/clip_grads.py — test-only PR

Checklist

• Commits signed off (git commit -s)
• Rebased on main
• Test-only change, no production code touched

Fixes #5827

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

This PR has been automatically converted to draft because all PRs must start as drafts.

When you are ready for review, click Ready for Review to begin the review process. This will:

1. Add the oncall reviewer (optional reviewer)
2. Add required review teams based on your changes

See the contribution guide for more details.

[dimapihtar]

/ok to test 254c131

[guihong-nv]

@Mattral please fix the failing tests.

[Mattral]

Thanks for the ping

I'll investigate the unit test failure (want to confirm whether it's related to this change or a pre-existing issue in the suite) and push a fix by Monday.

[Mattral]

@guihong-nv . Found the root cause: the test assumed a single rank, but CI runs unit tests with --nproc-per-node 8. Every rank seeds identically and computes the same local grads, so the SUM-based all_reduce (grad_stats_parallel_group=None → WORLD group) correctly multiplies the result by world_size, the inf-norm case uses MAX instead, which is why only [2.0] and [1.0] failed. Pushed a fix that accounts for world_size in the test's expected value, watching CI now to confirm.

[Mattral]

Hi @guihong-nv

following up here. The fix landed last week (test now accounts for world_size in the SUM vs MAX all-reduce comparison). I notice a couple of checks are still pending

let me know if those need an approval on your end, or if there's anything else you'd like adjusted before another look.

[guihong-nv]

@Mattral may you please resolve the conflict here?

[Mattral]

@guihong-nv
Resolved, kept both: my get_grad_norm_fp32 regression tests and your TestCountZerosFp32GtpPadding class, merged cleanly with no overlap in what they cover. Ready for another look.