​阿里通义千问：本地部署Qwen1.5开源大模型

一、Qwen1.5介绍

        通义千问为阿里云研发的大语言系列模型。千问模型基于Transformer架构，在超大规模的预训练数据上进行训练得到。预训练数据类型多样，覆盖广泛，包括大量网络文本、专业书籍、代码等。同时，在预训练模型的基础之上，使用对齐机制打造了模型的chat版本。

二、本地部署Qwen1.5

        由于资源问题，本机只能使用cpu运行qwen1.5_4b_chat，所以本次操作基于上述条件执行。

3.1 下载模型文件

可以Hugging Face Hub 下载模型，如果从 HuggingFace 下载比较慢，也可以从 ModelScope 中下载：

#从ModelScope下载模型
git clone https://www.modelscope.cn/qwen/Qwen1.5-4B-Chat.git

3.2 安装依赖

        可以通过 Conda 创建Python环境安装依赖。

# vi requirements.txt

# basic requirements

fastapi>=0.110.0
uvicorn>=0.29.0
pydantic>=2.7.0
tiktoken>=0.6.0
sse-starlette>=2.0.0

transformers>=4.37.0
torch>=2.1.0
sentencepiece>=0.2.0
sentence-transformers>=2.4.0
accelerate


#安装依赖
pip install requirements.txt

        安装过程可能重新网络问题，无法下载相关文件，可以重复执行多次；也可以去PyPI下载相关whl文件，离线安装。

3.2 编码实现

        通用python编码实现部署模型和提供RESTful API服务。

# 文件名：api-server-llm.py

# -*- coding: utf-8 -*-
# This is a sample Python script.

import argparse
import datetime
import uuid
from threading import Thread
import os

from typing import List, Optional
from pydantic import BaseModel
# Press Shift+F10 to execute it or replace it with your code.
# Press Double Shift to search everywhere for classes, files, tool windows, actions, and settings.
import uvicorn
from fastapi import FastAPI, Request, Response
from fastapi.responses import JSONResponse
from sse_starlette.sse import EventSourceResponse
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
from utils import num_tokens_from_string


# 声明API
app = FastAPI(default_timeout=1000 * 60 * 10)


# chat消息
class ChatMessage(BaseModel):
    # 角色
    role: str = None
    # chat内容
    content: str = None


# chat请求信息
class ChatRequest(BaseModel):
    # 对话信息
    messages: List[ChatMessage]
    # 模型平台
    platform: str = None
    # 具体模型名称
    model: str = None
    # 客户端id
    clientId: str = None
    stream: Optional[bool] = False


# 当前对话的模型输出内容
class ChatResponseChoice(BaseModel):
    # 模型推理终止的原因
    finishReason: str
    # 消息内容
    messages: ChatMessage


# tokens 统计数据
class ResponseUsage(BaseModel):
    # 模型输入的 tokens 数量
    outputTokens: int
    # 用户输入的 tokens 数量
    inputTokens: int


# chat返回信息
class ChatResponse(BaseModel):
    clientId: str = None
    usage: ResponseUsage
    choices: List[ChatResponseChoice]


# 统一异常处理
@app.exception_handler(Exception)
async def all_exceptions_handler(request: Request, exc: Exception):
    """
    处理所有异常
    """
    return JSONResponse(
        status_code=500,
        content={
            "msg": str(exc)
        }
    )


# 自定义中间件
@app.middleware("http")
async def unified_interception(request: Request, call_next):
    # 在这里编写你的拦截逻辑
    # 例如，检查请求的header或参数
    # 如果不满足条件，可以直接返回响应，不再调用后续的路由处理
    token = request.headers.get("Authorization")

    # if token is None:
    #     return JSONResponse({"message": "Missing Authorization"}, status_code=401)

    # 如果满足条件，则继续调用后续的路由处理
    response = await call_next(request)

    # 在这里编写你的响应处理逻辑
    # 例如，添加或修改响应头

    # 返回最终的响应
    return response


# 监控
@app.get("/api/models")
async def models():
    # 构造返回数据
    response = {
        "model": MODEL_NAME
    }
    return JSONResponse(response, status_code=200)


# 构造 回复消息对象
def create_chat_response(request_id: str, role: str, content: str, input_tokens: int, output_tokens: int,
                         finish_reason: str):
    now = datetime.datetime.now()  # 获取当前时间
    time = now.strftime("%Y-%m-%d %H:%M:%S")  # 格式化时间为字符串
    chat_message = {
        "output": {
            "choices": [{
                "finish_reason": finish_reason,
                "message": {
                    "role": role,
                    "content": content
                }
            }]
        },
        "request_id": request_id,
        "usage": {
            "input_tokens": input_tokens,
            "output_tokens": output_tokens,
            "total_tokens": (input_tokens + output_tokens)
        },
        "time": time
    }
    return chat_message


# 获取流式数据
def predict_stream(generation_kwargs: dict, request_id: str, input_tokens: int):
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()
    output_tokens = 0

    for new_text in streamer:
        output_tokens = output_tokens + num_tokens_from_string(str(new_text))
        # 构造返回数据
        yield str(create_chat_response(request_id=request_id, role="assistant", content=new_text,
                                       input_tokens=input_tokens, output_tokens=output_tokens, finish_reason=""))

    yield str(create_chat_response(request_id=request_id, role="assistant", content="", input_tokens=input_tokens,
                                   output_tokens=output_tokens, finish_reason="stop"))


@app.post("/api/v1/chat/completions")
async def completions(request: ChatRequest):
    # if MODEL_NAME != request.model:
    #     return JSONResponse({"msg": "模型异常！"}, status_code=500)

    if len(request.messages) <= 0:
        return JSONResponse({"msg": "消息不能为空！"}, status_code=500)

    # 请求id
    request_id = str(uuid.uuid4())
    # 计算输入tokens
    input_tokens = sum(num_tokens_from_string(message.content) for message in request.messages)

    # 使用自带的聊天模板，格式化对话记录
    text = tokenizer.apply_chat_template(request.messages, tokenize=False, add_generation_prompt=True)
    model_inputs = tokenizer(text, return_tensors="pt").to("cpu")

    if request.stream:
        print("----流式-----")
        # 流式对象

        generation_kwargs = dict(**model_inputs, streamer=streamer, max_new_tokens=4096,
                                 pad_token_id=tokenizer.eos_token_id,
                                 num_beams=1, do_sample=True, top_p=0.8,
                                 temperature=0.3)
        generate = predict_stream(generation_kwargs=generation_kwargs, request_id=request_id, input_tokens=input_tokens)

        return EventSourceResponse(generate, media_type="text/event-stream")

    print("----非流式-----")
    # 非流式
    generated_ids = model.generate(
        model_inputs.input_ids
        , max_new_tokens=512
    )
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
    # 计算输出tokens
    output_tokens = num_tokens_from_string(response)
    # 构造返回数据
    result = create_chat_response(request_id=request_id, role="assistant", content=response, input_tokens=input_tokens,
                                  output_tokens=output_tokens, finish_reason="stop")
    return JSONResponse(result, status_code=200)


if __name__ == '__main__':
    # 定义命令行解析器对象
    parser = argparse.ArgumentParser(description='大语言模型参数解析器')
    # 添加命令行参数、默认值
    parser.add_argument("--host", type=str, default="0.0.0.0")
    parser.add_argument("--port", type=int, default=8860)
    parser.add_argument("--model_path", type=str, default="")
    parser.add_argument("--model_name", type=str, default="")
    # 从命令行中结构化解析参数
    args = parser.parse_args()

    # 模型路径
    MODEL_PATH = args.model_path
    MODEL_NAME = args.model_name

    if len(MODEL_PATH) <= 0:
        raise Exception("大语言模型路径不能为空！")
    # 如果没有传入模型名称，则从路径中获取
    if len(MODEL_NAME) == 0:
       MODEL_DIR, MODEL_NAME = os.path.split(MODEL_PATH)

    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)

    model = AutoModelForCausalLM.from_pretrained(
        MODEL_PATH, trust_remote_code=True, device_map='auto'
    ).eval()
    # 流式输出
    streamer = TextIteratorStreamer(tokenizer=tokenizer, skip_prompt=True, skip_special_tokens=True)

    # 启动 uvicorn 服务
    uvicorn.run(app, host=args.host, port=args.port)

3.3 运行服务

# 运行脚本，指定模型路径
python api-server-llm.py --model_path=F:\llm\model\Qwen1.5-4B-Chat

3.4 使用API访问

curl -X POST -H "Content-Type: application/json" 'http://localhost:8860/api/v1/chat/completions' \
-d '{
	"platform": "tongyi", 
	"model": "Qwen1.5-4B-Chat",  
	"stream":true, 
	"messages":[ 
		{
			"role": "user", 
			"content": "你好，你是谁?" 
		} 
	]
  }'