Implementation:Vllm project Vllm Pydantic Schema Generation

Knowledge Sources	vLLM vLLM Docs
Domains	LLM Inference, Structured Output, Schema Definition
Last Updated	2026-02-08 13:00 GMT

Overview

Concrete pattern for defining output schemas (JSON Schema, regex, grammar, choice list) before constrained generation, provided by Pydantic and Python standard types.

Description

Before running constrained generation in vLLM, the user must define the desired output format. This implementation covers the four supported schema types and how to produce them:

• JSON Schema via Pydantic: Define a Pydantic BaseModel subclass with typed fields, then call model_json_schema() to produce a JSON Schema dictionary. Enums can be used for constrained field values.
• Regex: Supply a Python regex string that the entire output must match.
• GBNF Grammar: Supply a string in GBNF (GGML BNF) notation defining a context-free grammar.
• Choice List: Supply a list[str] of allowed output values.

Each schema type is passed into StructuredOutputsParams as the corresponding keyword argument (json, regex, grammar, or choice).

Usage

Use this pattern at the beginning of any structured output workflow. Define the schema in user code before constructing StructuredOutputsParams and SamplingParams.

Code Reference

Source Location

• Repository: vllm
• File: examples/offline_inference/structured_outputs.py

Signature

# JSON Schema via Pydantic
pydantic.BaseModel.model_json_schema() -> dict

# Regex pattern
pattern: str  # e.g., r"\w+@\w+\.com\n"

# GBNF grammar
grammar: str  # e.g., 'root ::= select_statement\n...'

# Choice list
choices: list[str]  # e.g., ["Positive", "Negative"]

Import

from pydantic import BaseModel
from enum import Enum

I/O Contract

Inputs

Name	Type	Required	Description
Pydantic model class	type[BaseModel]	No (use one schema type)	A Pydantic model whose model_json_schema() method produces the JSON Schema dict
regex	str	No (use one schema type)	A regular expression pattern the output must match
grammar	str	No (use one schema type)	A GBNF grammar string defining the output language
choice	list[str]	No (use one schema type)	A list of allowed output strings

Outputs

Name	Type	Description
json_schema	dict	JSON Schema dictionary produced by model_json_schema(), ready to pass to StructuredOutputsParams(json=...)
regex_pattern	str	Regex string ready to pass to StructuredOutputsParams(regex=...)
grammar_string	str	GBNF grammar string ready to pass to StructuredOutputsParams(grammar=...)
choice_list	list[str]	List of strings ready to pass to StructuredOutputsParams(choice=...)

Usage Examples

JSON Schema via Pydantic Model

from enum import Enum
from pydantic import BaseModel

class CarType(str, Enum):
    sedan = "sedan"
    suv = "SUV"
    truck = "Truck"
    coupe = "Coupe"

class CarDescription(BaseModel):
    brand: str
    model: str
    car_type: CarType

json_schema = CarDescription.model_json_schema()
# json_schema is a dict like:
# {'properties': {'brand': {'type': 'string'}, ...}, 'required': [...], ...}

Regex Pattern

email_pattern = r"\w+@\w+\.com\n"
# Pass to StructuredOutputsParams(regex=email_pattern)

GBNF Grammar

simplified_sql_grammar = """
root ::= select_statement
select_statement ::= "SELECT " column " from " table " where " condition
column ::= "col_1 " | "col_2 "
table ::= "table_1 " | "table_2 "
condition ::= column "= " number
number ::= "1 " | "2 "
"""
# Pass to StructuredOutputsParams(grammar=simplified_sql_grammar)

Choice List

choices = ["Positive", "Negative"]
# Pass to StructuredOutputsParams(choice=choices)

Related Pages

Implements Principle

• Principle:Vllm_project_Vllm_Output_Schema_Definition

Requires Environment

• Environment:Vllm_project_Vllm_Python_Dependencies