[megatron, recipe] fix: error of megatron init while detached actor and rollout

[lalala-2]

What does this PR do?

• We found that when asynchronously training and separately creating the Megatron's ActorRolloutRefWorker, the code incorrectly skipped the setting of ppo_mini_batch_size, and introduced an unnecessary mpu.initialize_model_parallel. These mistakes have imposed some constraints on the rollout resource configuration.
• Currently, the parameter synchronization of fully async and one-step-off does not take into account the situation of param offload, which is supported here as well.

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 correctly addresses an initialization error for detached Megatron actor/rollout workers by making model parallel initialization conditional. It also adds support for parameter offloading during weight synchronization, which is a valuable enhancement. The changes are logical and well-implemented. I have one suggestion to improve code consistency in recipe/fully_async_policy/megatron_worker.py.