[trainer, fully_async] feat: add streaming rollouter mode to the V1 PPO trainer

[huaiyizhao]

What does this PR do?

Adds a new fully_async trainer mode to the V1 PPO trainer that decouples rollout generation from the training step, so generation and training overlap instead of running lock-step (one batch fed per step). This brings the streaming-rollouter capability of verl/experimental/fully_async_policy to the V1 trainer, while reusing the existing TransferQueue as the streaming channel (no separate MessageQueue / Rollouter actor).

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

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 introduces a new fully_async (streaming) mode for the V1 PPO trainer, allowing an autonomous background feeder to continuously stream prompts into the TransferQueue under an in-flight budget. This decouples rollout generation from training and overlaps the two processes. The changes include the core streaming feeder logic, the PPOTrainerFullyAsync implementation, thread-safety guards for the training dataloader, configuration updates, and comprehensive CPU unit/integration tests and E2E scripts. The review feedback highlights two key improvement opportunities: first, updating the parameter version only during actual weight synchronization steps to prevent underestimating off-policy staleness; second, refactoring the feeder's pause/resume mechanism using a threading.Condition to eliminate polling delays and avoid cumulative GPU idle time.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Updating self._param_version at the end of every training step is incorrect because weight synchronization with the standalone rollout only occurs periodically (every parameter_sync_step steps). Tagging prompts with intermediate step numbers makes the trainer believe they were generated with fresher weights than they actually were, which underestimates off-policy staleness and can bypass the max_off_policy_threshold safety guard.

self._param_version should only be updated to self.global_steps when a weight synchronization actually occurs.

Suggested change

	with self._param_version_lock:
	self._param_version = self.global_steps
	if self.global_steps % self.parameter_sync_step == 0:
	with self._param_version_lock:
	self._param_version = self.global_steps

[gemini-code-assist]

Using self._paused.wait(self._poll_interval) or polling self._paused.is_set() with a sleep interval introduces an unnecessary delay (up to feeder_poll_interval, which defaults to 1.0s) when resuming the feeder after a weight synchronization. This can lead to significant cumulative GPU idle time during training.

Using a threading.Condition variable allows the feeder to be resumed instantly and efficiently without any polling or sleeping delays.

        self._stop = threading.Event()
        self._paused = False
        self._cv = threading.Condition()
        self.error = False  # set True if the feeder thread dies unexpectedly
        self._thread: threading.Thread | None = None

    def _loop(self):
        while not self._stop.is_set():
            with self._cv:
                while self._paused and not self._stop.is_set():
                    self._cv.wait()
            if self._stop.is_set():
                break
            try:
                counts = self._count_inflight()
                inflight = counts["pending"] + counts["running"] + counts["finished"] + counts["failure"]
                if inflight < self._budget:
                    self._feed_one_batch(self._param_version())
                else:
                    # interruptible sleep: avoids a tight busy loop while the budget is full
                    self._stop.wait(self._poll_interval)
            except StopIteration:
                logger.info("Streaming feeder: dataset exhausted, stopping feeder")
                break
            except Exception:
                logger.exception("Streaming feeder thread crashed")
                self.error = True
                break

    def start(self):
        """Start the background feeder thread."""
        self._stop.clear()
        with self._cv:
            self._paused = False
        self.error = False
        self._thread = threading.Thread(target=self._loop, name="streaming-rollout-feeder", daemon=True)
        self._thread.start()

    def stop(self, timeout: float = 30.0):
        """Signal the feeder to stop and join the thread."""
        self._stop.set()
        with self._cv:
            self._cv.notify_all()
        if self._thread is not None and self._thread.is_alive():
            self._thread.join(timeout=timeout)

    def pause(self):
        """Pause dispatching new prompts (e.g. during a weight sync).

        Generation already in flight is unaffected (it is aborted+continued by the checkpoint
        engine / FullyAsyncLLMServerClient). Idempotent.
        """
        with self._cv:
            self._paused = True

    def resume(self):
        """Resume dispatching after :meth:`pause`. Idempotent."""
        with self._cv:
            self._paused = False
            self._cv.notify_all()

    @property
    def paused(self) -> bool:
        with self._cv:
            return self._paused

[wuxibin89]

@huaiyizhao Thanks for you contribution. I have 2 questions:

1. What's the advantage of streaming rollouter mode compared to lock-step mode?
2. Do you have e2e experiment result on streaming rollouter mode?

[CLAassistant]

All committers have signed the CLA.

[huaiyizhao]

@wuxibin89 Hi.

1. Streaming mode keeps the rollout side as an autonomous capacity-driven service. It accepts input prompts whenever the inflight budget has room, which makes it more robust to rollout-side irregularities such as failures, drops, partial rollout abort/continue, long-tail requests, or uneven completion. In the ideal steady state where warmup exactly fills the budget and every step consumes/refills the same amount, its throughput can be close to warmup-based separate async. The practical benefit is flexibility and recovery under non-ideal rollout behavior.
2. No, it is an re-implementation of fully async trainer in v1. I expect similar speedups compared with colocated mode.
3. I have pushed another version with rollout-level agentloop worker, which also aims for a more fine-grained control over the rollout process.
4. The other trainers are left unchanged