[BREAKING] [rollout] feat: add a separate rollout worker

[vermouth1992]

What does this PR do?

• Introduce a separate rolloutworker that can be instantiated without hybridengine
• Introduce a ModelConfig that wraps all model related config
• Remove hf_rollout (will replace with TP support in the future if needed)
• Next PR: modify MegatronWorker to use separate rollout worker

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 significant and beneficial refactoring by creating a separate RolloutWorker and HFModelConfig. This change effectively decouples the rollout logic from the HybridEngine, improving modularity and maintainability. The implementation is well-executed, but I've identified one critical issue that will cause a runtime error. Please see the detailed comment below.

[vermouth1992]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces a separate RolloutWorker and a HFModelConfig to encapsulate model configurations, which is a positive step towards better modularity. However, the review has identified several critical issues in the new RolloutWorker implementation that must be addressed. These include a feature regression where hf rollout support has been dropped, a bug that will cause a crash due to using a configuration object before it's assigned, and another potential crash when accessing a processor object that can be None.

[gemini-code-assist]

self.config is used here before it is assigned on line 55. This will raise an AttributeError at runtime. You should use the config object passed as an argument to the __init__ method instead.

Suggested change

	timeout=datetime.timedelta(seconds=self.config.get("nccl_timeout", 600)),
	timeout=datetime.timedelta(seconds=config.get("nccl_timeout", 600)),

[gemini-code-assist]

The hf rollout type is not handled here, while vllm and sglang are. The previous implementation in fsdp_workers.py supported hf rollout. This appears to be an unintentional feature regression. If hf rollout is still intended to be supported, its implementation should be added here to avoid breaking existing functionality.

[gemini-code-assist]

self.model_config.processor can be None if a processor is not available for the model. This will cause an AttributeError when trying to access eos_token_id or pad_token_id. You should use self.model_config.get_processor() which safely returns either the processor or the tokenizer, preventing a potential crash.

Suggested change

	meta_info = {
	"eos_token_id": self.model_config.generation_config.eos_token_id
	if self.model_config.generation_config is not None
	else self.model_config.processor.eos_token_id,
	"pad_token_id": self.model_config.generation_config.pad_token_id
	if self.model_config.generation_config is not None
	else self.model_config.processor.pad_token_id,
	}
	processor = self.model_config.get_processor()
	meta_info = {
	"eos_token_id": self.model_config.generation_config.eos_token_id
	if self.model_config.generation_config is not None
	else processor.eos_token_id,
	"pad_token_id": self.model_config.generation_config.pad_token_id
	if self.model_config.generation_config is not None
	else processor.pad_token_id,
	}

[wuxibin89]

Should we separate different backend to subclass?

class vLLMRolloutWorker(RolloutWorker):
    pass

class SGLangRolloutWorker(RolloutWorker):
    pass