[chenjiaoAngel]

What does this PR do?

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

validate process and rollout process can be paralled in fully-async mode. This can reduce rollouter ilde and improve timing_s/gen speed when in short-long response_length case.

This is below profiler: (qwen2.5-Math-7b)

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, cfg, reward
  • 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.

we can use dapo_7b_math_fsdp2_8_8.sh, and add +async_training.parallel_validate_and_rollout=True command

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 parallel processing for validation in the FullyAsyncRollouter by using a ThreadPoolExecutor. This allows validation to run asynchronously, controlled by the parallel_validate_and_rollout configuration. The changes refactor the validation logic to support both synchronous and asynchronous execution paths. My review identifies a critical issue with unhandled exceptions in a fire-and-forget asyncio task, which could lead to silent failures. I have also pointed out several high-severity issues related to Python best practices, such as placing imports inside methods, accessing private attributes, using non-idiomatic boolean comparisons, and including non-English comments, all of which impact code quality and maintainability.

[bhks]

Why do we need more than 1 thread in this case ? Aren't we only making single validate call during validation ?

On high end GPU nodes there are > 100s of CPU core and this will essentially create 100s of unused threads and take memory 100*8MB.

[bhks]

Here we are only making call to validate method once.

[chenjiaoAngel]

This is subject to conditional restrictions, which is controlled by need_validate (consistent with the previous trigger conditions).

[bhks]

I was focusing more on the number of threads required to make a single validate call.

[chenjiaoAngel]

The number of thread control by generate. validate process only has one thread