[rollout, perf, cfg] fix: Add global step info and support more profile control params for rollout profiling (sglang backend) #5025

bithighrr · 2026-01-22T16:10:10Z

What does this PR do?

Address key functionality gaps in rollout discrete profiling for the sglang backend by adding global step awareness and expanding support for flexible profile control parameters:

Missing global step information resulted in disorganized profiling files;

Lack of support for specifying critical sglang backend profile control parameters (including num_steps, profile_by_stage, and merge_profiles) led to overly large profile files and hindered convenient profiling analysis.

Solutions implemented:

Pass kwargs parameters to the rollout server to enable global step awareness, and generate an independent folder for each global step to organize profiling data in a structured way;

Extend the profiling config by adding optional parameters to the content field, and remove unnecessary restrictions on content values in TorchProfilerToolConfig to support specifying sglang-specific profile control parameters (e.g., profile_by_stage, merge_profiles).

Checklist Before Starting

Search for similar PRs. Paste at least one query link here: PR #4320: [perf] feat: verl profiler system support Agent Loop scenario and integrate torch.profiler
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

To validate the functionality of each profile control parameter for the sglang rollout backend, orthogonal test cases were designed (minimizing redundant combinations while covering all key parameter values). The test results are as follows:

step_start	step_end	profile-by-stage	merge-profiles	stack	shapes	cpu/cuda	Test Result
default	default	default (False)	default (False)	default (False)	default (False)	default	Pass
0 (≥0)	1 (≥1)	default	default	default	default	default	Pass
0 (≥0)	default	default	default	default	default	default	Pass
default	1 (≥1)	default	default	default	default	default	Pass
default	default	True	default	default	default	default	Pass
default	default	default	True	default	default	default	Pass
default	default	default	default	True	default	default	Pass
default	default	default	default	default	True	default	Pass
default	default	default	default	default	default	[] / [cpu] / [cuda] / [cpu,cuda]	Pass

API and Usage Example

  rollout:
    quantization: null
    profiler:
      enable: True
      all_ranks: False
      ranks: [0, 1, 2] 
      tool_config:
        torch:
          discrete: True
          step_start: 0
          step_end: 1
          contents: [cpu, cuda, stack, shapes, profile_by_stage, merge_profiles]

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.

…le control params for rollout profiling (sglang backend)

gemini-code-assist

Code Review

This pull request introduces global step awareness for rollout profiling and expands support for sglang-specific profile control parameters. The changes correctly integrate global_step into the profiling process and add new configuration options for TorchProfilerToolConfig. However, there are a couple of high-severity issues related to the profiling logic in async_sglang_server.py that need to be addressed to ensure correct and predictable behavior.

gemini-code-assist · 2026-01-22T16:16:50Z

verl/workers/rollout/sglang_rollout/async_sglang_server.py

+        rollout_start_step = max(1, config.step_start or 1)
+        rollout_end_step = max(2, config.step_end or 2)


The hardcoded max(1, ...) and max(2, ...) for rollout_start_step and rollout_end_step can lead to unintended profiling ranges. For example, if a user sets config.step_start=0 and config.step_end=1 (intending to profile only step 0), rollout_end_step will be forced to 2, resulting in profiling steps 0 and 1. This overrides the user's explicit configuration. It's generally better to respect user-provided values or implement clearer default logic.

Suggested change

rollout_start_step = max(1, config.step_start or 1)

rollout_end_step = max(2, config.step_end or 2)

rollout_start_step = config.step_start if config.step_start is not None else 0

rollout_end_step = config.step_end if config.step_end is not None else -1

if rollout_end_step != -1 and rollout_end_step < rollout_start_step:

raise ValueError("rollout_end_step cannot be less than rollout_start_step")

For SGlang profiling, the forward step is required to start from 1.

gemini-code-assist · 2026-01-22T16:16:50Z

verl/workers/rollout/sglang_rollout/async_sglang_server.py

+        assert rollout_num_steps is None or rollout_num_steps > 0, (
+            f"Rollout num steps must be greater than 0 for sglang, but got {rollout_num_steps}"
+        )
+        self._auto_stop_profiling = rollout_num_steps is not None


The logic for _auto_stop_profiling seems inverted. self._auto_stop_profiling is set to True if rollout_num_steps is defined (i.e., config.step_start and config.step_end are provided). However, the stop_profile method's condition and not self._auto_stop_profiling means that if rollout_num_steps is set, the explicit self.tokenizer_manager.stop_profile() call will be skipped. This could prevent profiling sessions from being properly terminated, leading to resource leaks or corrupted profile files.

Suggested change

self._auto_stop_profiling = rollout_num_steps is not None

self._auto_stop_profiling = rollout_num_steps is not None and rollout_num_steps > 0

rollout_num_steps can only be a number here and will never get None, so _auto_stop_profiling is always True

Once the rollout_num_steps parameter is set, SGLang will automatically stop the profiling session.

This has been fixed.

tardis-key · 2026-01-23T01:30:28Z

verl/utils/profiler/config.py


    def __post_init__(self) -> None:
        """config validation logics go here"""
+        __support_contents = ["cuda", "cpu", "memory", "shapes", "stack", "profile-by-stage", "merge-profiles"]


Please explain the two newly added contents and try to find an appropriate place to update the documentation(at least add it here as a code comment)

The purpose of the parameters is explained in the _profile_args interface. This may be added to the documentation in future work.

tardis-key · 2026-01-23T01:38:53Z

verl/workers/rollout/sglang_rollout/async_sglang_server.py

+        assert rollout_num_steps is None or rollout_num_steps > 0, (
+            f"Rollout num steps must be greater than 0 for sglang, but got {rollout_num_steps}"
+        )
+        self._auto_stop_profiling = rollout_num_steps is not None


rollout_num_steps can only be a number here and will never get None, so _auto_stop_profiling is always True

tardis-key · 2026-01-23T01:43:52Z

verl/workers/rollout/sglang_rollout/async_sglang_server.py

            self.profiler_controller.check_enable()
            and self.profiler_controller.check_this_rank()
            and self.profiler_controller.is_discrete_mode()
+            and not self._auto_stop_profiling


If auto_stop_profiling is always True, it seems that stop_profile will never execute? Please verify the logic related to auto_stop_profiling

tardis-key · 2026-01-23T01:52:33Z

verl/workers/rollout/vllm_rollout/vllm_async_server.py


    async def start_profile(self, **kwargs):
+        # TODO: Persist global_step to engine server-created file/path
+        kwargs.pop("global_step")


keep the TODO, and remove the pop operation

An extra unused parameter should not cause any issue. However, if we need to use it later, we might spend time to figure out why global_step has vanished ^_^

The AsyncLLM.start_profile interface from vLLM does not support the global_step parameter ~

mengchengTang · 2026-01-23T03:36:13Z

verl/workers/rollout/sglang_rollout/async_sglang_server.py


        return TokenOutput(token_ids=token_ids, log_probs=log_probs, routed_experts=routed_experts)

+    def _profile_args(self, **kwargs) -> dict[str, Any]:


Keep Profiler Logic Cohesive: The profiler logic seems to be scattered across the main workflow. It would be better to encapsulate it to keep the main execution path clean and cohesive. Avoid adding too much profiling-specific code directly into the core rollout logic

This should be the SGLang implementation class, where potential differing interfaces should be reserved to allow for better encapsulation and abstraction design in the future.

I've pushed a new commit to improve code cohesion and reduce coupling. Thank you for your advice.

mengchengTang · 2026-01-23T03:37:07Z

verl/utils/profiler/config.py


    def __post_init__(self) -> None:
        """config validation logics go here"""
+        __support_contents = ["cuda", "cpu", "memory", "shapes", "stack", "profile-by-stage", "merge-profiles"]


Stick to Common Parameters: Some parameters added here seem specific to the SGLang interface (e.g., profile-by-stage). I suggest removing them if they don't provide significant benefits for profiling analysis. It's better to stick to system parameters that are common across both vLLM and SGLang backends to maintain consistency and reduce maintenance overhead.

This should include both generic parameters and profiling parameters specific to the rollout backend implementation, to prevent the inability to use the rollout backend's performance analysis capabilities later on.

[rollout, perf, cfg] fix: Add global step info and support more profi…

4d65cbd

…le control params for rollout profiling (sglang backend)

bithighrr requested review from PeterSH6, SwordFaith, chenhaiq, eric-haibin-lin, tongyx361, vermouth1992, wuxibin89 and zhaochenyang20 as code owners January 22, 2026 16:10

gemini-code-assist bot reviewed Jan 22, 2026

View reviewed changes

tardis-key requested changes Jan 23, 2026

View reviewed changes

mengchengTang reviewed Jan 23, 2026

View reviewed changes

bithighrr added 2 commits January 23, 2026 03:51

Prevented _auto_stop_profiling from being erroneously set to True

907b37e

Improved code cohesion and reduced coupling.

5e8e82c

bithighrr force-pushed the 0121 branch from 3bb7713 to 5e8e82c Compare January 23, 2026 07:43

		rollout_start_step = max(1, config.step_start or 1)
		rollout_end_step = max(2, config.step_end or 2)

-        rollout_start_step = max(1, config.step_start or 1)
-        rollout_end_step = max(2, config.step_end or 2)
+        rollout_start_step = config.step_start if config.step_start is not None else 0
+        rollout_end_step = config.step_end if config.step_end is not None else -1
+        if rollout_end_step != -1 and rollout_end_step < rollout_start_step:
+            raise ValueError("rollout_end_step cannot be less than rollout_start_step")

	self._auto_stop_profiling = rollout_num_steps is not None
	self._auto_stop_profiling = rollout_num_steps is not None and rollout_num_steps > 0


		return TokenOutput(token_ids=token_ids, log_probs=log_probs, routed_experts=routed_experts)

		def _profile_args(self, **kwargs) -> dict[str, Any]:

[rollout, perf, cfg] fix: Add global step info and support more profile control params for rollout profiling (sglang backend) #5025

Are you sure you want to change the base?

[rollout, perf, cfg] fix: Add global step info and support more profile control params for rollout profiling (sglang backend) #5025

Conversation

bithighrr commented Jan 22, 2026

What does this PR do?

Checklist Before Starting

Test

API and Usage Example

Design & Code Changes

Checklist Before Submitting

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist bot Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants