[megatron] fix: mbridge load optimizer dist_ckpt

[ccilery]

What does this PR do?

fix training resume from mbridge load optimizer dist_ckpt.

This PR fix #3517

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)

[gemini-code-assist]

Code Review

This pull request fixes an issue with resuming training from a distributed optimizer checkpoint loaded via mbridge. The changes introduce an is_loading flag to differentiate behavior in generate_state_dict when loading checkpoints, which is a good approach. However, there is a concern about modifying torch.serialization's global state at module import time, which can lead to side effects. I've suggested moving this logic to an explicit initialization point.

[gemini-code-assist]

Modifying global state at module import time is generally discouraged as it can lead to hard-to-debug side effects and makes the system's behavior dependent on import order. Here, torch.serialization._safe_globals is modified when this module is imported.

This logic should be moved into an explicit initialization function that is called once during your application's setup, for example at the beginning of the load_checkpoint method. This makes the dependency explicit and avoids unintended consequences in other parts of the codebase that might also use torch.serialization.

[ISEEKYAN]

@ccilery can you fix the pre-commit formatter

[ccilery]

@ccilery can you fix the pre-commit formatter

@ISEEKYAN Thanks, have fixed.