[megatron, ckpt] refactor: optimize mcore ckpt manager impl

[ETOgaosion]

What does this PR do?

Refactoring Megatron checkpoint menager to avoid complex logics, following the way mentioned in #5630

Problem

Now the mcore checkpoint manager has multiple backends: mbridge, dist_checkpointing, verl original implemenation, and there are 2 model related save_contents APIs: model, hf_model, which is complex in implemenation.

Also, the API implementation in codes is complex, save_contents are transformed with should_save_xxx judgement and the backends converted to use_xxx judgement. These naming is confusing and makes users difficult to understand saving logics.

Refactor

So, let's make things clearer for mcore checkpoint mechanism.

First, refer to FSDP checkpoint APIs, model now means save sharded models, and hf_model will save the whole model.

To align with FSDP, mcore uses the same API.

For mcore, there is no sharded models, and have dist_checkpoint and mbridge backends, so in theory:

• When mbridge enabled, model and hf_model should work the same, they all save huggingface model (only save once)
• When mbridge disabled and use dist_checkpoint, model should use mcore dist_checkpoint implementation and hf_model should save hf_model use verl's implementation (old methods, should deprecate to reduce maintenance burden)

For future and checkpoint manager refactor:

• mbridge shall be enabled by default
• dist_checkpoint as a fall back method in background when model in save_contents
  • We can leave it to be an API default as false
  • We will not support both mbridge and dist_checkpoint as false
• hf_model should only supported by mbridge

Shown in Figure:

cc @ISEEKYAN @vermouth1992 @PeterSH6 @wuxibin89 @khazic @tongyx361

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, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • 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).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

This pull request refactors MegatronCheckpointManager to support modular checkpointing backends, primarily introducing use_mbridge for HuggingFace format weights. The update replaces monolithic state dict generation with composable builders to avoid unnecessary overhead and includes a new suite of unit tests. Technical feedback recommends using a try...finally block to ensure transformer configuration attributes are restored after serialization and removing redundant type checks for callable objects.