Sglang

Overview

SGLang is a fast LLM inference and serving framework with structured generation (constrained decoding), RadixAttention for prefix caching, and an OpenAI-compatible API. Particularly strong for reasoning models (DeepSeek R1, QwQ) and guided generation with JSON schema, grammar, and regex constraints.

Installation

uv pip install sglang[all]

Quick Start

import sglang as sgl

@sgl.function
def multi_turn(s, question):
    s += sgl.system("You are a helpful assistant.")
    s += sgl.user(question)
    s += sgl.assistant()

state = multi_turn.run(question="What is the derivative of x^2?")
print(state["answer"])

Server Mode

python -m sglang.launch_server --model-path Qwen/Qwen2.5-1.5B-Instruct --port 30000

from openai import OpenAI
client = OpenAI(base_url="http://localhost:30000/v1", api_key="none")

Structured Generation

@sgl.function
def json_gen(s):
    s += "Generate a person's info in JSON."
    s += sgl.gen("json_output", max_tokens=128, 
                 schema='{"type": "object", "properties": {"name": {"type": "string"}, "age": {"type": "integer"}}}')

state = json_gen.run()
print(state["json_output"])

References

• SGLang docs
• SGLang GitHub