Environment:Volcengine Verl SGLang Rollout Environment

sources: Repo|verl|https://github.com/volcengine/verl

domains: Infrastructure, Inference

last_updated: 2026-02-07 17:00 GMT

Overview

Optional SGLang environment for high-performance rollout generation with multi-turn tool use support in verl.

Description

verl uses SGLang as an alternative rollout backend, especially for multi-turn agent loop scenarios. SGLang 0.5.6 is the pinned version in setup.py. For async server mode, sglang >= 0.5.5 is required. SGLang requires torch==2.9.1 when installed via the sglang extra.

Usage

Required when using SGLang as the rollout engine (rollout.name=sglang in config), especially for multi-turn tool-calling workflows.

System Requirements

NVIDIA GPU
Linux

Dependencies

sglang[srt,openai] == 0.5.6
torch == 2.9.1 (when using sglang extra)

Quick Install

pip install "verl[sglang]"

or

pip install "sglang[srt,openai]==0.5.6"

Code Evidence

From verl/utils/import_utils.py:45-51:

@cache
def is_sglang_available():
    try:
        sglang_spec = importlib.util.find_spec("sglang")
    except ModuleNotFoundError:
        sglang_spec = None
    return sglang_spec is not None

And from verl/workers/rollout/sglang_rollout/async_sglang_server.py:158:

assert version.parse(sglang.__version__) >= version.parse("0.5.5")

Common Errors

"sglang not found" -> pip install sglang
"sglang version must be >= 0.5.5" -> Upgrade sglang
torch version mismatch -> Install torch==2.9.1

Compatibility Notes

SGLang and vLLM cannot be used simultaneously as rollout backends. SGLang has FP8 block quantization support for performance optimization.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment