Fix mxfp8 param gather numerical issue when DP overlap is off

[WanZzzzzz]

What does this PR do ?

main: #4800

With MXFP8 grad-buffer reuse and non-overlap param gather, each DistOpt stages its own updated main-param shards into its param buffers during step. However, param sync is a DDP model-chunk operation: model_chunk.start_param_sync() gathers both dense and expert bucket groups, copies gathered values into model weights, and zeros the shared MXFP8 param/grad buffers. For MoE, dense and expert DistOpts may share the same model chunk, so defer param sync until all chained optimizers have staged their params, then sync each model chunk once. Otherwise the expert DistOpt may copy the zeroes into expert model weights.

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share a design doc with the team. If you're unsure what's the best way to do so, contact the @mcore-oncall.

Issue tracking

For PRs from open-source community contributors:

• New features: a linked issue is required. Please open a feature request and reference it here before submitting the PR.
• Small updates (bug fixes, minor improvements): a linked issue is recommended and will accelerate the PR review process.

Linked issue:

Contribution process

Pre-checks

• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh on my PR

Code review

Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.

Step 1: Mark PR as "Ready for Review"

1. When your PR is ready, click Ready for Review.
2. An oncall reviewer is auto-assigned and expert reviewers are notified based on your changes.
   • Some PRs may jump straight to step 2. This is determined by .github/CODEOWNERS.

⚠️ Only mark as ready once merge-conflicts are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

Step 2: Final Review

For PRs that change megatron/core, once all expert reviewers have approved, the Final Review label is applied automatically and final reviewers are assigned.

For PRs outside megatron/core, this step is skipped.

Step 3: Approved

Once all required reviewers have approved, the Approved label is applied automatically.

Merge

Any member of mcore-engineers will be able to merge your PR.

For MRs into `dev` branch

The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

[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.

[WanZzzzzz]

@zhongbozhu @kunlunl

[zhongbozhu]

LGTM, cc @kunlunl

[kunlunl]

I understand the cause of this bug, and I think the fix is correct. However, is there a way we can refactor the code to make it less hacky? For example, if we can pass information via function args rather than using ._attribute?

[WanZzzzzz]

I understand the cause of this bug, and I think the fix is correct. However, is there a way we can refactor the code to make it less hacky? For example, if we can pass information via function args rather than using ._attribute?

Yeah I considered adding defer_param_sync to step_with_ready_grads(), but that method is part of the common MegatronOptimizer interface and is implemented by multiple optimizer wrappers. The flag is only meaningful for DistributedOptimizer’s non-overlap param-gather path. Adding this flag to the method would expand the public optimizer API for a very specific issue which I personally don't like that.

[kunlunl]

/ok to test d68550d

[kunlunl]

@WanZzzzzz Could you fix the CI?

@Victarry Do you think we can merge this one now?

[WanZzzzzz]

/ok to test 4dc3559

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/26072551439

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/26083859497