Expose `engine_kwargs` from SGLang to Verl configuration by hebiao064 · Pull Request #1616 · verl-project/verl

hebiao064 · 2025-05-21T06:44:16Z

Checklist Before Starting

Search for similar PR(s).

What does this PR do?

Expose engine_kwargs from SGLang to Verl configuration

This PR enables RL users to configure engine_kwargs directly through Verl, providing more control and flexibility over inference behavior.

High-Level Design

One key motivation is the choice of attention backend, which can significantly affect rollout performance. The SGLang team has observed that different attention backends perform better in different phases:

FA3 tends to be more efficient during the prefill stage.

FlashInfer or Triton generally offer better performance during decode.

Moreover, the optimal backend may change across versions of SGLang. By exposing these parameters, we allow users to tune their setup based on the specific use case and version, ultimately improving performance and adaptability.

Add one-line overview of what this PR aims to achieve or accomplish.

Demonstrate the high-level design if this PR is complex.

Specific Changes

List the specific changes.

API

Demonstrate how the API changes if any.

Usage Example

Provide usage example(s) for easier usage.

# Add code snippet or script demonstrating how to use this

Test

In my setup about QWen 2.5 7B Instruct on H200:

timing_s/step:106.761 (flash infer)
timing_s/step:100.520 （fa3)
timing_s/step:100.364 (triton)

Hence, I would suggest our team to use fa3 or triton for now.

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, evaluatuion results, etc.

Additional Info.

Issue Number: Fixes issue # or discussion # if any.
Training: [Note which backend this PR will affect: FSDP, Megatron, both, or none]
Inference: [Note which backend this PR will affect: vLLM, SGLang, both, or none]

Checklist Before Submitting

Read the Contribute Guide.
Apply pre-commit checks.
Add [BREAKING] to the PR title if it breaks any API.
Update the documentation about your changes in the docs.
Add CI test(s) if necessary.

zhaochenyang20 · 2025-05-21T06:48:53Z

great work. I think for VLM, FA3 could be even better?

SwordFaith · 2025-05-21T06:49:34Z

Which backend is best as default ?

zhaochenyang20 · 2025-05-21T06:50:40Z

Which backend is best as default ?

right now any of it is okay. I think fa3 is default?

hebiao064 · 2025-05-21T07:00:20Z

Which backend is best as default ?

It's determined by SGLang's logic, for now, it's fa3 on hopper, flashinfer on other arch.

https://github.com/sgl-project/sglang/blob/4024e1d2a853903cfddfb70f5e10e165db30305c/python/sglang/srt/utils.py#L2096

But even on hopper or ampere, sometimes we found triton is faster, that's why we need this granularity.

hebiao064 · 2025-05-21T07:16:45Z

great work. I think for VLM, FA3 could be even better?

It depends on the batch size, context size and response length, at least we can provide the freedom to user to decide.

vermouth1992 · 2025-05-21T13:59:04Z

@zhaochenyang20 Shall we merge this?

zhaochenyang20 · 2025-05-21T15:49:14Z

@vermouth1992 thanks, merged

…t#1616) ### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? Expose `engine_kwargs` from SGLang to Verl configuration This PR enables RL users to configure `engine_kwargs` directly through Verl, providing more control and flexibility over inference behavior. ### High-Level Design One key motivation is the choice of attention backend, which can significantly affect rollout performance. The SGLang team has observed that different attention backends perform better in different phases: FA3 tends to be more efficient during the prefill stage. FlashInfer or Triton generally offer better performance during decode. Moreover, the optimal backend may change across versions of SGLang. By exposing these parameters, we allow users to tune their setup based on the specific use case and version, ultimately improving performance and adaptability. > Add one-line overview of what this PR aims to achieve or accomplish. > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API > Demonstrate how the API changes if any. ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### Test In my setup about QWen 2.5 7B Instruct on H200: ``` timing_s/step:106.761 (flash infer) timing_s/step:100.520 （fa3) timing_s/step:100.364 (triton) ``` Hence, I would suggest our team to use fa3 or triton for now. > 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? #1616 causes vllm engine arg init failed, not know why CI of that PR fail to detect. Some errors have shown up. ![image](https://github.com/user-attachments/assets/ac6bb86e-1576-458e-b341-0e949724ac12) We may better separate engine args for different inference systems ### High-Level Design > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API ```yml engine_kwargs: # inference engine parameters vllm: swap_space: null # null means "use the engine default value" (usually 4 GB), setting it to, e.g., 32 means 32 GB sglang: attention_backend: null # null means use the engine default value, available options: flashinfer, triton, flashmla ``` ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? verl-project#1616 causes vllm engine arg init failed, not know why CI of that PR fail to detect. Some errors have shown up. ![image](https://github.com/user-attachments/assets/ac6bb86e-1576-458e-b341-0e949724ac12) We may better separate engine args for different inference systems ### High-Level Design > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API ```yml engine_kwargs: # inference engine parameters vllm: swap_space: null # null means "use the engine default value" (usually 4 GB), setting it to, e.g., 32 means 32 GB sglang: attention_backend: null # null means use the engine default value, available options: flashinfer, triton, flashmla ``` ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

…t#1616) ### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? Expose `engine_kwargs` from SGLang to Verl configuration This PR enables RL users to configure `engine_kwargs` directly through Verl, providing more control and flexibility over inference behavior. ### High-Level Design One key motivation is the choice of attention backend, which can significantly affect rollout performance. The SGLang team has observed that different attention backends perform better in different phases: FA3 tends to be more efficient during the prefill stage. FlashInfer or Triton generally offer better performance during decode. Moreover, the optimal backend may change across versions of SGLang. By exposing these parameters, we allow users to tune their setup based on the specific use case and version, ultimately improving performance and adaptability. > Add one-line overview of what this PR aims to achieve or accomplish. > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API > Demonstrate how the API changes if any. ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### Test In my setup about QWen 2.5 7B Instruct on H200: ``` timing_s/step:106.761 (flash infer) timing_s/step:100.520 （fa3) timing_s/step:100.364 (triton) ``` Hence, I would suggest our team to use fa3 or triton for now. > 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? verl-project#1616 causes vllm engine arg init failed, not know why CI of that PR fail to detect. Some errors have shown up. ![image](https://github.com/user-attachments/assets/ac6bb86e-1576-458e-b341-0e949724ac12) We may better separate engine args for different inference systems ### High-Level Design > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API ```yml engine_kwargs: # inference engine parameters vllm: swap_space: null # null means "use the engine default value" (usually 4 GB), setting it to, e.g., 32 means 32 GB sglang: attention_backend: null # null means use the engine default value, available options: flashinfer, triton, flashmla ``` ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

…t#1616) ### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? Expose `engine_kwargs` from SGLang to Verl configuration This PR enables RL users to configure `engine_kwargs` directly through Verl, providing more control and flexibility over inference behavior. ### High-Level Design One key motivation is the choice of attention backend, which can significantly affect rollout performance. The SGLang team has observed that different attention backends perform better in different phases: FA3 tends to be more efficient during the prefill stage. FlashInfer or Triton generally offer better performance during decode. Moreover, the optimal backend may change across versions of SGLang. By exposing these parameters, we allow users to tune their setup based on the specific use case and version, ultimately improving performance and adaptability. > Add one-line overview of what this PR aims to achieve or accomplish. > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API > Demonstrate how the API changes if any. ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### Test In my setup about QWen 2.5 7B Instruct on H200: ``` timing_s/step:106.761 (flash infer) timing_s/step:100.520 （fa3) timing_s/step:100.364 (triton) ``` Hence, I would suggest our team to use fa3 or triton for now. > 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? verl-project#1616 causes vllm engine arg init failed, not know why CI of that PR fail to detect. Some errors have shown up. ![image](https://github.com/user-attachments/assets/ac6bb86e-1576-458e-b341-0e949724ac12) We may better separate engine args for different inference systems ### High-Level Design > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API ```yml engine_kwargs: # inference engine parameters vllm: swap_space: null # null means "use the engine default value" (usually 4 GB), setting it to, e.g., 32 means 32 GB sglang: attention_backend: null # null means use the engine default value, available options: flashinfer, triton, flashmla ``` ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

…t#1616) ### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? Expose `engine_kwargs` from SGLang to Verl configuration This PR enables RL users to configure `engine_kwargs` directly through Verl, providing more control and flexibility over inference behavior. ### High-Level Design One key motivation is the choice of attention backend, which can significantly affect rollout performance. The SGLang team has observed that different attention backends perform better in different phases: FA3 tends to be more efficient during the prefill stage. FlashInfer or Triton generally offer better performance during decode. Moreover, the optimal backend may change across versions of SGLang. By exposing these parameters, we allow users to tune their setup based on the specific use case and version, ultimately improving performance and adaptability. > Add one-line overview of what this PR aims to achieve or accomplish. > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API > Demonstrate how the API changes if any. ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### Test In my setup about QWen 2.5 7B Instruct on H200: ``` timing_s/step:106.761 (flash infer) timing_s/step:100.520 （fa3) timing_s/step:100.364 (triton) ``` Hence, I would suggest our team to use fa3 or triton for now. > 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

### Checklist Before Starting - [x] Search for similar PR(s). ### What does this PR do? verl-project#1616 causes vllm engine arg init failed, not know why CI of that PR fail to detect. Some errors have shown up. ![image](https://github.com/user-attachments/assets/ac6bb86e-1576-458e-b341-0e949724ac12) We may better separate engine args for different inference systems ### High-Level Design > Demonstrate the high-level design if this PR is complex. ### Specific Changes > List the specific changes. ### API ```yml engine_kwargs: # inference engine parameters vllm: swap_space: null # null means "use the engine default value" (usually 4 GB), setting it to, e.g., 32 means 32 GB sglang: attention_backend: null # null means use the engine default value, available options: flashinfer, triton, flashmla ``` ### Usage Example > Provide usage example(s) for easier usage. ```python # Add code snippet or script demonstrating how to use this ``` ### 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, evaluatuion results, etc. ### Additional Info. - **Issue Number**: Fixes issue # or discussion # if any. - **Training**: [Note which backend this PR will affect: FSDP, Megatron, both, or none] - **Inference**: [Note which backend this PR will affect: vLLM, SGLang, both, or none] ### Checklist Before Submitting - [ ] Read the [Contribute Guide](https://github.com/volcengine/verl?tab=readme-ov-file#contribution-guide). - [ ] Apply [pre-commit checks](https://github.com/volcengine/verl?tab=readme-ov-file#code-linting-and-formatting). - [ ] Add `[BREAKING]` to the PR title if it breaks any API. - [ ] Update the documentation about your changes in the [docs](https://github.com/volcengine/verl/tree/main/docs). - [ ] Add CI test(s) if necessary.

Support Flexible Engine Kwargs for SGLang

5b21301

set default to null

ea019b3

zhaochenyang20 approved these changes May 21, 2025

View reviewed changes

zhaochenyang20 merged commit 72683a7 into verl-project:main May 21, 2025
33 checks passed

ETOgaosion mentioned this pull request May 22, 2025

[BugFix] Fix sglang and vllm engine args #1634

Merged

6 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Expose `engine_kwargs` from SGLang to Verl configuration#1616

Expose `engine_kwargs` from SGLang to Verl configuration#1616
zhaochenyang20 merged 2 commits intoverl-project:mainfrom
hebiao064:support_eng_args_for_sglang

hebiao064 commented May 21, 2025

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

SwordFaith commented May 21, 2025

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

hebiao064 commented May 21, 2025 •

edited

Loading

Uh oh!

hebiao064 commented May 21, 2025

Uh oh!

vermouth1992 commented May 21, 2025

Uh oh!

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

hebiao064 commented May 21, 2025

Checklist Before Starting

What does this PR do?

High-Level Design

Specific Changes

API

Usage Example

Test

Additional Info.

Checklist Before Submitting

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

SwordFaith commented May 21, 2025

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

hebiao064 commented May 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hebiao064 commented May 21, 2025

Uh oh!

vermouth1992 commented May 21, 2025

Uh oh!

Uh oh!

zhaochenyang20 commented May 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

hebiao064 commented May 21, 2025 •

edited

Loading