跳到内容

结构化输出

vLLM 支持使用 xgrammarguidance 作为后端来生成结构化输出。本文档将向您展示一些可用选项的示例,以生成结构化输出。

警告

如果您仍在使用以下在 v0.12.0 中移除的已弃用的 API 字段,请更新您的代码以使用 structured_outputs,具体方法请参见本文档其余部分。

  • guided_json -> {"structured_outputs": {"json": ...}}StructuredOutputsParams(json=...)
  • guided_regex -> {"structured_outputs": {"regex": ...}}StructuredOutputsParams(regex=...)
  • guided_choice -> {"structured_outputs": {"choice": ...}}StructuredOutputsParams(choice=...)
  • guided_grammar -> {"structured_outputs": {"grammar": ...}}StructuredOutputsParams(grammar=...)
  • guided_whitespace_pattern -> {"structured_outputs": {"whitespace_pattern": ...}}StructuredOutputsParams(whitespace_pattern=...)
  • structural_tag -> {"structured_outputs": {"structural_tag": ...}}StructuredOutputsParams(structural_tag=...)
  • guided_decoding_backend -> 从您的请求中移除此字段

在线服务 (OpenAI API)

您可以使用 OpenAI 的 CompletionsChat API 来生成结构化输出。

支持以下参数,必须添加为额外参数:

  • choice:输出将是选项中的一个。
  • regex:输出将遵循正则表达式模式。
  • json:输出将遵循 JSON 模式。
  • grammar:输出将遵循上下文无关文法。
  • structural_tag:在生成的文本中,遵循指定标签集内的 JSON 模式。

您可以在 OpenAI 兼容服务器 页面上找到支持的参数的完整列表。

结构化输出在 OpenAI 兼容服务器中默认支持。您可以通过将 --structured-outputs-config.backend 标志设置为 vllm serve 来选择要使用的后端。默认后端是 auto,它将尝试根据请求的详细信息选择合适的后端。您也可以选择一个特定的后端,以及一些选项。完整的选项集可在 vllm serve --help 文本中找到。

现在让我们从最简单的 choice 开始,看每个用例的示例。

代码
from openai import OpenAI
client = OpenAI(
    base_url="https://:8000/v1",
    api_key="-",
)
model = client.models.list().data[0].id

completion = client.chat.completions.create(
    model=model,
    messages=[
        {"role": "user", "content": "Classify this sentiment: vLLM is wonderful!"}
    ],
    extra_body={"structured_outputs": {"choice": ["positive", "negative"]}},
)
print(completion.choices[0].message.content)

下一个示例展示了如何使用 regex。支持的正则表达式语法取决于结构化输出后端。例如,xgrammarguidanceoutlines 使用 Rust 风格的正则表达式,而 lm-format-enforcer 使用 Python 的 re 模块。目的是在给定一个简单的正则表达式模板的情况下,生成一个电子邮件地址。

代码
completion = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": "Generate an example email address for Alan Turing, who works in Enigma. End in .com and new line. Example result: [email protected]\n",
        }
    ],
    extra_body={"structured_outputs": {"regex": r"\w+@\w+\.com\n"}, "stop": ["\n"]},
)
print(completion.choices[0].message.content)

结构化文本生成中最相关的特性之一是生成具有预定义字段和格式的有效 JSON 的选项。为此,我们可以通过两种方式使用 json 参数:

下一个示例展示了如何将 response_format 参数与 Pydantic 模型一起使用。

代码
from pydantic import BaseModel
from enum import Enum

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()

completion = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": "Generate a JSON with the brand, model and car_type of the most iconic car from the 90's",
        }
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "car-description",
            "schema": CarDescription.model_json_schema()
        },
    },
)
print(completion.choices[0].message.content)

提示

虽然并非严格必需,但通常最好在提示中指明 JSON 模式以及如何填充字段。在大多数情况下,这可以显著提高结果。

最后是 grammar 选项,这可能是最难使用的,但它非常强大。它允许我们定义完整的语言,如 SQL 查询。它通过使用上下文无关的 EBNF 文法来实现。例如,我们可以使用它来定义一种特定格式的简化 SQL 查询。

代码
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 "
"""

completion = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": "Generate an SQL query to show the 'username' and 'email' from the 'users' table.",
        }
    ],
    extra_body={"structured_outputs": {"grammar": simplified_sql_grammar}},
)
print(completion.choices[0].message.content)

另请参见:完整示例

推理输出

您也可以使用结构化输出来进行推理,支持:用于推理模型。

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --reasoning-parser deepseek_r1

请注意,您可以将推理与任何提供的结构化输出功能结合使用。以下示例使用了 JSON 模式:

代码
from pydantic import BaseModel


class People(BaseModel):
    name: str
    age: int


completion = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": "Generate a JSON with the name and age of one random person.",
        }
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "people",
            "schema": People.model_json_schema()
        }
    },
)
print("reasoning: ", completion.choices[0].message.reasoning)
print("content: ", completion.choices[0].message.content)

另请参见:完整示例

实验性自动解析 (OpenAI API)

本节介绍 OpenAI 的 Beta 版包装器,它覆盖了 client.chat.completions.create() 方法,并提供了与 Python 特定类型更丰富的集成。

在撰写本文时(openai==1.54.4),这是 OpenAI 客户端库中的一个“Beta”功能。代码参考可以在此处找到。

在以下示例中,vLLM 使用 vllm serve meta-llama/Llama-3.1-8B-Instruct 进行设置。

这是一个简单的示例,演示了如何使用 Pydantic 模型获取结构化输出。

代码
from pydantic import BaseModel
from openai import OpenAI

class Info(BaseModel):
    name: str
    age: int

client = OpenAI(base_url="http://0.0.0.0:8000/v1", api_key="dummy")
model = client.models.list().data[0].id
completion = client.beta.chat.completions.parse(
    model=model,
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "My name is Cameron, I'm 28. What's my name and age?"},
    ],
    response_format=Info,
)

message = completion.choices[0].message
print(message)
assert message.parsed
print("Name:", message.parsed.name)
print("Age:", message.parsed.age)
ParsedChatCompletionMessage[Testing](content='{"name": "Cameron", "age": 28}', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[], parsed=Testing(name='Cameron', age=28))
Name: Cameron
Age: 28

这是一个更复杂的示例,使用嵌套的 Pydantic 模型来处理分步数学解决方案。

代码
from typing import List
from pydantic import BaseModel
from openai import OpenAI

class Step(BaseModel):
    explanation: str
    output: str

class MathResponse(BaseModel):
    steps: list[Step]
    final_answer: str

completion = client.beta.chat.completions.parse(
    model=model,
    messages=[
        {"role": "system", "content": "You are a helpful expert math tutor."},
        {"role": "user", "content": "Solve 8x + 31 = 2."},
    ],
    response_format=MathResponse,
)

message = completion.choices[0].message
print(message)
assert message.parsed
for i, step in enumerate(message.parsed.steps):
    print(f"Step #{i}:", step)
print("Answer:", message.parsed.final_answer)

输出

ParsedChatCompletionMessage[MathResponse](content='{ "steps": [{ "explanation": "First, let\'s isolate the term with the variable \'x\'. To do this, we\'ll subtract 31 from both sides of the equation.", "output": "8x + 31 - 31 = 2 - 31"}, { "explanation": "By subtracting 31 from both sides, we simplify the equation to 8x = -29.", "output": "8x = -29"}, { "explanation": "Next, let\'s isolate \'x\' by dividing both sides of the equation by 8.", "output": "8x / 8 = -29 / 8"}], "final_answer": "x = -29/8" }', refusal=None, role='assistant', audio=None, function_call=None, tool_calls=[], parsed=MathResponse(steps=[Step(explanation="First, let's isolate the term with the variable 'x'. To do this, we'll subtract 31 from both sides of the equation.", output='8x + 31 - 31 = 2 - 31'), Step(explanation='By subtracting 31 from both sides, we simplify the equation to 8x = -29.', output='8x = -29'), Step(explanation="Next, let's isolate 'x' by dividing both sides of the equation by 8.", output='8x / 8 = -29 / 8')], final_answer='x = -29/8'))
Step #0: explanation="First, let's isolate the term with the variable 'x'. To do this, we'll subtract 31 from both sides of the equation." output='8x + 31 - 31 = 2 - 31'
Step #1: explanation='By subtracting 31 from both sides, we simplify the equation to 8x = -29.' output='8x = -29'
Step #2: explanation="Next, let's isolate 'x' by dividing both sides of the equation by 8." output='8x / 8 = -29 / 8'
Answer: x = -29/8

关于使用 structural_tag 的示例可以在此处找到: examples/online_serving/structured_outputs

离线推理

离线推理支持相同类型的结构化输出。要使用它,我们需要在 SamplingParams 中使用 StructuredOutputsParams 类来配置结构化输出。StructuredOutputsParams 内的主要可用选项是:

  • json
  • regex
  • choice
  • grammar
  • structural_tag

这些参数的使用方式与在线服务示例中的参数相同。下面展示了一个使用 choice 参数的示例。

代码
from vllm import LLM, SamplingParams
from vllm.sampling_params import StructuredOutputsParams

llm = LLM(model="HuggingFaceTB/SmolLM2-1.7B-Instruct")

structured_outputs_params = StructuredOutputsParams(choice=["Positive", "Negative"])
sampling_params = SamplingParams(structured_outputs=structured_outputs_params)
outputs = llm.generate(
    prompts="Classify this sentiment: vLLM is wonderful!",
    sampling_params=sampling_params,
)
print(outputs[0].outputs[0].text)

另请参见:完整示例