[perf] fix: profiler bug fix for agent loop

[mengchengTang]

What does this PR do?

Updated Description: This PR fixes the missing rollout profiling data issue caused by the Agent Loop architecture shift.

Root Cause: Inference requests now bypass the generate_sequences method of the worker, causing missing performance data during the rollout phase in Discrete Mode.

Solution: We add explicit profiling control to support the current rollout workers. This is a backward-compatible fix.
New methods start_capture_profile and stop_capture_profile in DistProfiler and Workers.
AgentLoopManager now manually starts and stops profiling via vLLMReplica around inference requests.

This ensures existing rollout workers can collect performance data correctly.

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).)

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces profiling capabilities for the agent loop by adding start_profile and stop_profile methods to AgentLoopManager, vLLMHttpServerBase, and vLLMReplica. The changes correctly propagate the profiling calls down to the workers. My review includes one high-severity suggestion to fix a blocking call (ray.get) within an asynchronous actor in vLLMHttpServerBase, which should be converted to a non-blocking asyncio.gather to prevent blocking the event loop and to maintain consistency with other asynchronous methods in the class.

[tardis-key]

It can be seen that the overall process control and the actual profiling startup are far apart, and the original solution requires passing the calling function many times, which is not elegant enough.

[tardis-key]

prof_step and profile_step are confusing. And new variables are unnecessary, the original role&profile_step can still meet the requirements.

[mengchengTang]

The corresponding variable(s) have been deprecated in the new solution.

[tardis-key]

@FightingZhen @wuxibin89 ready for review. This pr is bugfix for legacy_workers.

[wuxibin89]

This is only suitable for vllm with vLLMAsyncRollout(ModelRunner) colocated with trainer in same process, while sglang have separate processes. We're working on separating vllm process as well #4280.

For rollout performance profile, we should reuse vllm/sglang server metrics

• https://docs.vllm.ai/en/latest/design/metrics/
• https://docs.sglang.io/references/production_metrics.html

[wuxibin89]

Also see: https://verl.readthedocs.io/en/latest/advance/grafana_prometheus.html

[mengchengTang]

Understood. I completely agree that using native server metrics is the right direction.
And this pr is necessary as a bugfix for the currently supported architecture (colocated vLLM) to ensure profiling data is captured correctly right now. We treat this as an interim solution and will align with the native metrics approach after the process separation (#4280) lands.

[tardis-key]

Is there any way to unify start_e2e_profiler and start_capture_profiler?

[tardis-key]

what about moving start_async_rollout_profile to DistProfilerExtension?

[tardis-key]

Try to encapsulate the profiler method of rollout, and support custom use of either the engine's built-in profiler capability or the VeriL profiler capability.

[tardis-key]

1. Do not rename start/stop. The implementation of the second pair of functions depends on the patch implementation.
2. Patch the start_profiler capability of vLLM. Do not unravel vLLM's complex scheduling mechanism, so as to keep the external scheduling logic of VeRL cleaner.
3. Retain distinct serer profiling capabilities for SGLang and vLLM, and consider implementing rolloutExtension or slangExtension/vllmExtension.