[feat]: verl integrate msprobe data collection

[Tjh-UKN]

What does this PR do?

Integrates msprobe PrecisionDebugger into VERL’s rollout/ref/train/update paths with minimal, explicit start/stop calls. Dumps are organized as {data_dir}/{global_step}/{stage} for consistent step/stage separation. This adds a new precision_debugger config block and wires global_steps into training batches. (No related issues/PRs.)

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
  • 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

Not covered by CI. This change depends on the external msprobe runtime and requires a real training run to validate dump outputs. Please run a short PPO/GRPO step on FSDP and Megatron backends with precision_debugger.enable=true and confirm dump directories are created per stage.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Example config snippet (ppo_trainer.yaml / ppo_megatron_trainer.yaml)
precision_debugger:
  enable: True
  config_path: /path/to/config.json
  data_dir: outputs/precision_debug
  steps: [1, 2, 5]  # optional
  stages: ["rollout", "train_fwd", "train_bwd", "update_actor", "ref_model"]

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

• Design: Minimal, explicit start/stop around key stages; PrecisionDebugger.step() is always called before stop() to advance internal step counts. Dumps are organized by {global_step}/{stage}.
• Key changes:
  • Add precision_debugger config block to trainer configs and generated configs.
  • Inject global_steps into DataProto.meta_info in trainer loop.
  • Add precision_start/precision_stop helper (msprobe only) with staged dump paths.
  • Hook start/stop into:
    • Rollout (generate_sequences)
    • Ref model (compute_ref_log_prob)
    • Train forward/backward/update actor

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: msprobe runtime is external; this integration requires environment-specific profiling runs.
• 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.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

TAJh seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[wuxibin89]

Please move precision_start and precision_stop into DistProfiler

[Tjh-UKN]

I’ll do that

[wuxibin89]

1. Move it under global_profiler
2. Add a typed config class for it.
3. Add doc

[Tjh-UKN]

I‘ll move it
Here is the draft version[step1] and will be reviewed by profiler mates to decide the final version[step2]
Thanks for your tips

[tardis-key]

Due to the addition of the msprobe dependency introduced by this pr, it is necessary to update the corresponding requirements files, supplement the necessary CI , and update the Dockerfile.

[tardis-key]

Integrate into DistProfile control logic as much as possible to maximize reuse and minimize modification. Maybe we can confirm the control flow design for this pr first.

[tardis-key]

Necessary test for different hardware platforms/training engines/rollout engines