Implementation:Sgl project Sglang Sgl Gen Constrained

Knowledge Sources	SGLang
Domains	NLP, Structured_Generation, Frontend_DSL
Last Updated	2026-02-10 00:00 GMT

Overview

Concrete tool for generating constrained text using regex or JSON schema patterns in the SGLang frontend DSL.

Description

The sgl.gen function creates an SglGen expression that, when executed within an @sgl.function, generates text optionally constrained by a regex pattern or json_schema string. For the Engine API, constraints are passed via the regex or json_schema keys in the sampling parameters dict. The constraint backends (outlines or xgrammar) handle the FSM-guided token masking.

Usage

Use sgl.gen(regex=...) in the frontend DSL or sampling_params={"regex": ...} in the Engine API to produce constrained output. Use sgl.gen(json_schema=...) for JSON output conforming to a schema.

Code Reference

Source Location

• Repository: sglang
• File: python/sglang/lang/api.py
• Lines: L75-139

Signature

def gen(
    name: Optional[str] = None,
    max_tokens: Optional[int] = None,
    min_tokens: Optional[int] = None,
    stop: Optional[Union[str, List[str]]] = None,
    temperature: Optional[float] = None,
    top_p: Optional[float] = None,
    top_k: Optional[int] = None,
    regex: Optional[str] = None,
    json_schema: Optional[str] = None,
    # ... additional sampling params
) -> SglGen:
    """Generate text with optional constraints."""

Import

import sglang as sgl

# Use within @sgl.function
s += sgl.gen("output", regex=r"\d+", max_tokens=10)

I/O Contract

Inputs

Name	Type	Required	Description
name	Optional[str]	No	Variable name to store the generated text
regex	Optional[str]	No	Regex pattern to constrain output
json_schema	Optional[str]	No	JSON schema string to constrain output
max_tokens	Optional[int]	No	Maximum tokens to generate
temperature	Optional[float]	No	Sampling temperature

Outputs

Name	Type	Description
SglGen	SglGen	Expression node for the frontend DSL execution graph

Usage Examples

Regex-Constrained Generation (Frontend DSL)

import sglang as sgl

@sgl.function
def extract_date(s, text):
    s += "Extract the date from this text: " + text + "\n"
    s += "Date: " + sgl.gen("date", regex=r"\d{4}-\d{2}-\d{2}", max_tokens=10)

state = extract_date.run(text="The event is on 2024-03-15.")
print(state["date"])  # "2024-03-15"

JSON Schema Generation (Frontend DSL)

import json
import sglang as sgl

schema = json.dumps({
    "type": "object",
    "properties": {
        "name": {"type": "string"},
        "age": {"type": "integer"},
    },
    "required": ["name", "age"],
})

@sgl.function
def extract_person(s, description):
    s += description + "\nExtract person info as JSON:\n"
    s += sgl.gen("json_output", json_schema=schema, max_tokens=128)

state = extract_person.run(description="John is 30 years old.")
person = json.loads(state["json_output"])

Engine API Constrained Generation

import sglang as sgl

engine = sgl.Engine(model_path="meta-llama/Llama-3.1-8B-Instruct")
output = engine.generate(
    "Generate a valid email address:",
    {"regex": r"[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}", "max_new_tokens": 32},
)
print(output["text"])  # Guaranteed valid email format

Related Pages

Implements Principle

• Principle:Sgl_project_Sglang_Constrained_Text_Generation

Requires Environment

• Environment:Sgl_project_Sglang_Python_Dependencies