[rollout] feat: add rollout config

[vermouth1992]

What does this PR do?

• Add rollout config

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 introduces a structured RolloutConfig using dataclasses to replace the previous dictionary-based configuration for rollouts. This is a great improvement for type safety and maintainability. The changes are consistently applied across the configuration files and their usage in fsdp_workers.py. My review focuses on the structure of the new RolloutConfig dataclass. I've identified two areas for improvement to make the configuration more robust and logically organized: encapsulating training sampling parameters and relocating actor-specific log probability parameters.

[gemini-code-assist]

The sampling parameters temperature, top_k, top_p, do_sample, and n are duplicated from SamplingConfig. This can lead to inconsistencies and makes the configuration harder to manage. As noted in the TODO on line 115, these should be encapsulated within a SamplingConfig object for training rollouts, similar to how val_kwargs is handled for validation. This will improve the structure and maintainability of the configuration. I recommend replacing these individual parameters with a train_kwargs field of type SamplingConfig. This will also require updating the corresponding YAML configuration files.

Suggested change

	temperature: float = 1.0
	top_k: int = -1
	top_p: float = 1.0
	do_sample: bool = True
	n: int = 1
	train_kwargs: SamplingConfig = field(default_factory=SamplingConfig)

[gemini-code-assist]

As the comment on line 123 correctly points out, the log probability computation is an actor responsibility. Therefore, the configuration parameters log_prob_micro_batch_size, log_prob_micro_batch_size_per_gpu, log_prob_use_dynamic_bsz, and log_prob_max_token_len_per_gpu should be part of ActorConfig, not RolloutConfig. Keeping them here is confusing and structurally incorrect, even if the values are linked from the actor config in YAML. Moving these parameters to ActorConfig would make the configuration more logical and easier to maintain.

[wuxibin89]

Should we move rollout sampling params to SamplingConfig?

[vermouth1992]

Yes. I left a TODO. Currently, we don't want to introduce breaking changes. RolloutConfig should match rollout.yaml exactly

[vermouth1992]

We also want to remove the logprob bsz into actor.infer in the following PRs