[trainer,algo] feat: Support On-Policy Distillation in main_ppo_sync

[0oshowero0]

What does this PR do?

Support #5041 in main_ppo_sync.py

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 implements teacher model distillation support in the PPO synchronization loop, including the addition of a TeacherModelManager and updates to the agent loop and transfer queue utilities to handle teacher logprobs and KVBatchMeta. Review feedback highlights a potential KeyError in agent_loop.py when extra_fields is missing and a crash in main_ppo_sync.py when processing multiple agent loop outputs.

[gemini-code-assist]

The code assumes extra_fields is always present in the output dictionary. However, since output is created using self.model_dump(exclude_unset=True) at line 212, if extra_fields was not explicitly set (it has a default empty dict), it will be excluded from the resulting dictionary. This will cause a KeyError when attempting to access output["extra_fields"]. It is safer to use .get() or ensure the key exists before popping.

        extra_fields = output.get("extra_fields", {})
        teacher_ids, teacher_logprobs = (
            extra_fields.pop("teacher_ids", None),
            extra_fields.pop("teacher_logprobs", None),
        )

[gemini-code-assist]

The code at lines 377-382 will crash if output is a list of AgentLoopOutput objects, as it attempts to access .prompt_ids and .response_ids directly on the output variable. While there is a TODO comment at line 376, the current implementation is broken for the multi-output case which is explicitly handled earlier in the function (line 350). You should iterate over the outputs list to compute teacher logprobs for each item.

Suggested change

	await self._compute_teacher_logprobs(
	output,
	prompt_ids=output.prompt_ids,
	response_ids=output.response_ids,
	validate=validate,
	)
	# TODO: Support output:list[AgentLoopOutput]
	for out in outputs:
	await self._compute_teacher_logprobs(
	out,
	prompt_ids=out.prompt_ids,
	response_ids=out.response_ids,
	validate=validate,
	)