[sglang] feat: add native sgl server

[ChangyiYang]

What does this PR do?

Summary

This PR introduces a native HTTP server implementation for SGLang, aiming to fundamentally improve flexibility, scalability, and integration capabilities. By transitioning to a more robust client-server architecture, this change addresses several core bottlenecks in the current design.

Key Changes

• Engine Replacement – Replaced the original sgl.Engine instance with a native HTTP server. ✅ Completed
• Distributed Optimization – Utilizing a server-based architecture to remove the requirement of gathering all data to TP rank 0. This change resolves the previous dist.barrier timeout issue by replacing the collective wait with per-sample synchronization. 🚧 In Progress
• Router Integration – Plan to integrate with the native SGLang router for streamlined request handling. 💡 Nice to have

Motivation

The current sgl.Engine driver model presents several architectural challenges, particularly in complex distributed environments. Moving to an HTTP server architecture is motivated by the need to solve the following critical issues:

1. Eliminate Data Flow Bottlenecks and Improve Performance:

   • Problem: The data flow logic of the existing driver process is misaligned with the training data flow. It requires all data for a single SGLang instance to be gathered to TP rank 0. This data is then processed by the tokenizer manager and sent via ZMQ to the various schedulers. As a result, the preprocess and postprocess steps are slower than expected.
   • Solution: The HTTP server architecture decentralizes this process, allowing each rank to handle requests independently. This removes the "gather to rank 0" bottleneck, dramatically improving data throughput and overall performance.

2. Resolve CPU Resource Contention:

   • Problem: At the request level, the SGLang driver object cannot be pickled for use in subprocesses. This limitation means that the request-level asynchronous rollout logic and the engine itself are forced to compete for the same CPU time slices, leading to performance degradation.
   • Solution: By decoupling the request handling (client) from the inference engine (server), we isolate the processes, eliminating the CPU contention and allowing for more efficient resource utilization.

3. Fix Distributed Synchronization Timeouts:

   • Problem: The dist.barrier timeout is a frequent issue where worker ranks remain idle while waiting for TP rank 0 to complete its intensive processing. This collective wait time creates inefficiency and can lead to failures.
   • Solution: The HTTP server model shifts this from a collective barrier to a per-sample synchronization. Workers communicate with the server as needed, removing the long wait times and making the distributed setup more stable and efficient.

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).)

[gemini-code-assist]

Code Review

This pull request introduces a native HTTP server for SGLang, which is a significant architectural improvement for flexibility and scalability. The implementation is comprehensive, including synchronous and asynchronous adapters, extensive test coverage, and example configurations. However, I've identified a few critical issues that need to be addressed. There's a configuration mismatch that will cause runtime errors, and the server launch logic for multi-node setups appears to be flawed, which would prevent it from working correctly in distributed environments. Additionally, there's a performance-related regression due to a hardcoded logging level.

[gemini-code-assist]

This refactoring has introduced a regression where engine_kwargs from the configuration are no longer passed to the SGLang engine constructor (AsyncEngine or AsyncHttpServerAdapter). This prevents overriding engine parameters like log_level. Additionally, log_level is now hardcoded to "info", which, as noted in the comments, can cause significant performance degradation and is not suitable for production. The engine_kwargs should be merged into the args dictionary, and the log_level should be configurable rather than hardcoded.

            args = {
                "model_path": actor_module,
                "dtype": self.config.dtype,
                "mem_fraction_static": self.config.gpu_memory_utilization,
                "enable_memory_saver": True,
                "base_gpu_id": 0,
                "gpu_id_step": 1,
                "tp_size": self._tp_size,
                "node_rank": node_rank,
                "load_format": load_format,
                "dist_init_addr": dist_init_addr,
                "nnodes": nnodes,
                "trust_remote_code": trust_remote_code,
                # NOTE(linjunrong): add rank to prevent SGLang generate same port inside PortArgs.init_new
                # when random.seed is being set during training
                "port": 30000 + rank,
                # NOTE(Chenyang): if you want to debug the SGLang engine output
                # please set the following parameters
                # Otherwise, it will make the engine run too slow
                # "log_level": "error",
                # log_requests=True,
                # log_requests_level=2,
                # NOTE(Chenyang): turn on max_running_requests to set the max concurrent running requests
                # max_running_requests=1,
                "mm_attention_backend": "fa3",
                "attention_backend": attention_backend if attention_backend is not None else "fa3",
                # In async mode, we want token in token out.
                "skip_tokenizer_init": self.config.mode == "async",
                **engine_kwargs,
            }

[zhaochenyang20]

FAILED test_sglang_async_rollout_w_interaction.py::test_async_sglang_rollout_w_interaction - omegaconf.errors.ConfigAttributeError: Missing key sglang_engine_mode
    full_key: sglang_engine_mode
    object_type=dict

@ChangyiYang fix the CI plz.

4c3310d

[zhaochenyang20]

https://github.com/volcengine/verl/actions/runs/17015367333/job/48239555678?pr=3090#step:12:813

@ChangyiYang rebase with main to solve this

[ChangyiYang]

@zhaochenyang20 some ci should be fixed now. Can you rerun it?

[zhaochenyang20]

https://github.com/volcengine/verl/actions/runs/17114945140/job/48721055839?pr=3090#step:7:321

maybe we need to register it

[zhaochenyang20]

Add a comment to say this would use the server.

[wuxibin89]

1. This PR is mainly addressing issues in batch mode with generate_sequences, while batch mode is going to be deprecated and switch to server mode by default. [rollout] feat: change rollout default mode from spmd to server mode #3161. So we tend to not make any new major change to batch mode except bugfix.
2. For the server mode, I found it's pretty easy to add native sglang server with a few lines:
   wuxibin89/verl@wuxibin/refactor_rollout2...wuxibin/sglang_native_server

[ChangyiYang]

Correct me if I am wrong.
In my opinion, the code @wuxibin89 provide is NOT providing the native sgl engine. We want the "generate“ api to send a http request a sgl server. But in the provided code, it launchs a sgl server but the generate api still goes to the local engine instance, which cannot allow the addition of sgl router in the future. So I think the provided code cannot do what this pr do. At lease, the sgl client that rollout holds (which send request to a sgl server) needs to provide the api that local engine provide

[ChangyiYang]

Thanks @zhaochenyang20 for detailed review and instruction! Thanks @SwordFaith for providing base code and instruction along the way!

[sangnekim]

Hello. thank you for your work.
I think the variable should be replaced sglang_engine_mode with sglang_rollout_mode.

• dist.barrier() timeout error issue still in progress?

[ChangyiYang]

Hi @sangnekim . Yes, the naming is better to change in your way. The dist barrier won't go in process since Verl would adopt agent loop and all the old rollout function would be deprecated.

[lizipao]

sglang_engine_mode

@ChangyiYang
TypeError: RolloutConfig.init() got an unexpected keyword argument 'sglang_rollout_mode'
Is this a bug?

[ChangyiYang]

@lizipao Yes, the config has a typo. I will change that. For now you can change this line

verl/verl/workers/config/rollout.py

Line 169 in 33edd95

sglang_engine_mode: str = "local"

[lizipao]

@lizipao Yes, the config has a typo. I will change that. For now you can change this line

verl/verl/workers/config/rollout.py

Line 169 in 33edd95

sglang_engine_mode: str = "local"

Have you encountered this error before?
[torch_memory_saver.cpp] CUresult error: 1 (invalid argument) file=csrc/core.cpp func=pause line=77
[http_server_engine.py:691] : Async request to release_memory_occupation timed out (attempt 1)

[ChangyiYang]

@lizipao I never encountered this error before.