[trainer, recipe] feat: Fully Async Policy add Rollout Importance Sampling with Megatron

[lalala-2]

What does this PR do?

Related to #3955 #2981
This PR support computing prox log prob with megatron in Fully Async Policy for Rollout Importance Sampling, and provided two example run scripts, supporting asynchronous RL training with GRPO for Qwen3-30B-A3B-Base.

Additional Notes: We found that the example script examples/grpo_trainer/run_qwen3moe-30b_megatron_96gb.sh originally used Qwen3-30B-A3B instead of Qwen3-30B-A3B-Base. Through our experimental validation, we observed significant train-inference mismatch issues when using Qwen3-30B-A3B for RL training (similar to the problem described in this GitHub issue).The issue is avoided with Qwen3-30B-A3B-Base, so we modified this example script to prevent others from encountering the same pitfall.

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 adds support for Rollout Importance Sampling in the Fully Async Policy with Megatron, introducing utilities for model parameter management between CPU and GPU. The changes are logical and well-structured, including new example run scripts. My review focuses on improving the new utility functions for clarity and performance, and correcting minor errors in the new shell scripts. I've identified an unused variable, an inconsistent use of pin_memory() that could impact performance, and duplicated code in the new shell scripts.

[szrlee]

@lalala-2 Thank you for your contribution! Could you take a look to the new system on rollout correction (rollout_is) from #3984 and migrate to this system?

the config document: https://verl.readthedocs.io/en/latest/advance/rollout_corr.html