support online mtp sft

[liuzhenhai93]

What does this PR do ?

support mtp sft during rl train
⚠️ 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:

• support mtp sft during rl train #4549

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.

[Victarry]

/ok to test e43c4d5

[ericharper]

Why do we need to do this in GPTModel? Can it happen in the RL loop?

[liuzhenhai93]

Why do we need to do this in GPTModel? Can it happen in the RL loop?

The gradient flow of mtp should be decoupled from the main model during rl training, which cannot be trivially achieved without intrusive modifications to Megatron core

[Victarry]

Thanks for the original proposal here. I refactored this into a new PR: #5080

The new version keeps the same online RL + MTP goal, but changes the API shape a bit:

• return_logits=True lets RL callers pass sampled labels for the MTP auxiliary CE loss while still receiving main LM logits for the external RL loss.
• mtp_isolated_loss=True isolates the MTP auxiliary objective from the main decoder hidden states, shared embeddings, and output layer weights.
• The default pretraining/SFT behavior remains unchanged.

I also added @liuzhenhai93 as a co-author on the new PR commit, since it is based on the original idea and implementation from this PR.