Windows 平台零基础部署 Qwen1.5 大模型教程

3. 运行 conda clean -i 清除索引缓存，保证用的是镜像站提供的索引。

注：后续如果想切换回官方源，.condarc 文件里面的内容清掉即可。

创建虚拟环境

1. 通过以下命令来创建虚拟环境：

conda create --name llm python=3.11.5

[图片：创建环境]

在这里我们创建的虚拟环境名称为 llm，选择 Python 的版本选择为 3.11.5。

2. 查看环境是否创建成功

通过命令：

conda info --envs

[图片：环境列表]

可以看到，我们创建的 llm 已经在里面了。

3. 接下来我们切换到创建的环境中

conda activate llm

[图片：激活环境]

部署大模型

下载大模型

以我们要部署的 Qwen1.5-0.5B-Chat 为例，通过 ModelScope 提供的下载方式，我们选择 Git 方式下载。

git clone https://www.modelscope.cn/qwen/Qwen1.5-0.5B-Chat.git

我把大模型下载到 D:\大模型\Qwen1.5-0.5B-Chat 目录下。

安装依赖

在这之前，我们先查一下我们的 cuda 版本。前面已经装好了相应的 NVIDIA 驱动，我们通过桌面右键，选择'NVIDIA 控制面板'->左下角'系统信息'->切换为'组件'Tab：

[图片：CUDA 版本查询]

可以看到我的 cuda 版本为 12.2。

安装 PyTorch

通过 PyTorch 官网找到相应的安装脚本。

这里选择 PyTorch 2.1.0 版本，之所以不选择更加新的版本，是因为会遇到其它问题。而 cuda 则选择 12.1 版本的，跟电脑的 12.2 比较接近。

conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda=12.1 -c pytorch -c nvidia

安装 transformers

我们需要安装 transformers 库：

conda install conda-forge::transformers

配置 VSCode

在写代码前，先配置一下写代码工具，那是必须的。这里以 VSCode 为例。

1. 下载安装

通过 VSCode 官网进行下载安装。

2. 安装拓展

[图片：插件市场]

同时推荐安装 autoDocstring - Python Docstring Generator、autopep8。这样 VSCode 在编写 Python 代码时，可以提高代码质量和编写效率：

• 快速生成代码的注释，提高代码可读性与可维护性。
• autopep8，代码格式化工具。

3. 引入 conda 环境

首先打开引入我们的大模型目录（D:\大模型）

[图片：VSCode 打开目录]

接着按住 Ctrl + Shift + P 打开命令面板，搜索'python 选择解析器'，选择上我们刚刚新建的 llm 环境。这样在 VSCode 中，我们就按照那个环境来跑我们的代码了。

[图片：选择解释器]

尝试跑 demo

1. 根据官方文档，首先在大模型同级目录下（D:\大模型），新建一个 qwen.py 文件，写入：

from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

# Now you do not need to add "trust_remote_code=True"
model = AutoModelForCausalLM.from_pretrained(
    "Qwen1.5-0.5B-Chat",  # 修改大模型位置
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen1.5-0.5B-Chat") # 修改大模型位置

# Instead of using model.chat(), we directly use model.generate()
# But you need to use tokenizer.apply_chat_template() to format your inputs as shown below
# 改成中文提问
prompt = "给我简单介绍一下大型语言模型。"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

# Directly use generate() and tokenizer.decode() to get the output.
# Use `max_new_tokens` to control the maximum output length.
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]

# 打印一下助手回复的内容
print(response)

注：这里要修改对大模型的引入路径。

2. 执行 py 文件：

python qwen.py

[图片：运行结果]

发现还是缺了一个依赖，安装：

conda install conda-forge::accelerate

再重新执行，这时候就可以看到模型的回复了。

[图片：模型回复]

如果想在 VSCode 中运行，则需要选择在'在专用终端中运行 Python 文件'。

[图片：VSCode 运行设置]

运行效果如下：

[图片：终端输出]

如果想要流式的，可以改为以下代码：

from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread

device = "cuda"  # the device to load the model onto

# Now you do not need to add "trust_remote_code=True"
model = AutoModelForCausalLM.from_pretrained(
    "Qwen1.5-0.5B-Chat",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Qwen1.5-0.5B-Chat")

# Instead of using model.chat(), we directly use model.generate()
# But you need to use tokenizer.apply_chat_template() to format your inputs as shown below
# 改成中文提问
prompt = "给我简单介绍一下大型语言模型。"
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

# Directly use generate() and tokenizer.decode() to get the output.
# Use `max_new_tokens` to control the maximum output length.
streamer = TextIteratorStreamer(
    tokenizer, skip_prompt=True, skip_special_tokens=True)
generation_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=512)
thread = Thread(target=model.generate, kwargs=generation_kwargs)

thread.start()
generated_text = ""
for new_text in streamer:
    generated_text += new_text
    print(generated_text)

部署为 API

这时候我们借助一个 FastAPI 和 Uvicorn 来实现 API 接口的支持。

• FastAPI： 是一个用于构建 API 的现代、快速（高性能）的 web 框架，使用 Python 并基于标准的 Python 类型提示。
• Uvicorn： 是一个快速的 ASGI（Asynchronous Server Gateway Interface）服务器，用于构建异步 Web 服务。

快速启动

1. 首先我们安装这两个库：

conda install fastapi uvicorn

2. 新建 web.py ，写入代码：

import uvicorn
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from argparse import ArgumentParser

app = FastAPI()

# 支持 CORS
app.add_middleware(
    CORSMiddleware,
    allow_origins=['*'],
    allow_credentials=True,
    allow_methods=['*'],
    allow_headers=['*'],
)


@app.get("/")
async def index():

    return {"message": "Hello World"}


def _get_args():
    parser = ArgumentParser()

    parser.add_argument('--server-port',
                        type=int,
                        default=8000,
                        help='Demo server port.')
    parser.add_argument('--server-name',
                        type=str,
                        default='127.0.0.1',
                        help='Demo server name. Default: 127.0.0.1, which is only visible from the local computer.'
                        ' If you want other computers to access your server, use 0.0.0.0 instead.',
                        )

    args = parser.parse_args()
    return args


if __name__ == '__main__':
    args = _get_args()

    uvicorn.run(app, host=args.server_name, port=args.server_port, workers=1)

3. 运行 web.py

python web.py

[图片：服务启动]

这时候请求接口，就可以看到返回 hello world ：

[图片：接口测试]

接入大模型测试

上面已经启动 API 服务了，我们把前面 Qwen 的代码写进去：

from contextlib import asynccontextmanager
import torch
import uvicorn
import time
from transformers import AutoModelForCausalLM, AutoTokenizer
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from argparse import ArgumentParser
from typing import List, Literal, Optional, Union
from pydantic import BaseModel, Field


@asynccontextmanager
async def lifespan(app: FastAPI):  # collects GPU memory
    yield
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.ipc_collect()

app = FastAPI(lifespan=lifespan)

# 支持 CORS
app.add_middleware(
    CORSMiddleware,
    allow_origins=['*'],
    allow_credentials=True,
    allow_methods=['*'],
    allow_headers=['*'],
)


class ChatMessage(BaseModel):
    role: Literal['user', 'assistant', 'system']
    content: Optional[str]


class DeltaMessage(BaseModel):
    role: Optional[Literal['user', 'assistant', 'system']] = None
    content: Optional[str] = None


class ChatCompletionRequest(BaseModel):
    model: str
    messages: List[ChatMessage]
    stream: Optional[bool] = False


class ChatCompletionResponseChoice(BaseModel):
    index: int
    message: ChatMessage
    finish_reason: Literal['stop', 'length']


class ChatCompletionResponseStreamChoice(BaseModel):
    index: int
    delta: DeltaMessage
    finish_reason: Optional[Literal['stop', 'length']]


class ChatCompletionResponse(BaseModel):
    model: str
    object: Literal['chat.completion', 'chat.completion.chunk']
    choices: List[Union[ChatCompletionResponseChoice,
                        ChatCompletionResponseStreamChoice]]
    created: Optional[int] = Field(default_factory=lambda: int(time.time()))


@app.get("/")
async def index():

    return {"message": "Hello World"}


@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
async def create_chat_completion(request: ChatCompletionRequest):
    global model, tokenizer

    # 简单的错误校验
    if request.messages[-1].role != "user":
        raise HTTPException(status_code=400, detail="Invalid request")

    text = tokenizer.apply_chat_template(
        request.messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    # Directly use generate() and tokenizer.decode() to get the output.
    # Use `max_new_tokens` to control the maximum output length.
    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]

    choice_data = ChatCompletionResponseChoice(
        index=0,
        message=ChatMessage(role="assistant", content=response),
        finish_reason="stop"
    )

    return ChatCompletionResponse(model=request.model, choices=[choice_data], object="chat.completion")


def _get_args():
    parser = ArgumentParser()

    parser.add_argument(
        '-c',
        '--checkpoint-path',
        type=str,
        default='Qwen1.5-0.5B-Chat',
        help='Checkpoint name or path, default to %(default)r',
    )

    parser.add_argument('--server-port',
                        type=int,
                        default=8000,
                        help='Demo server port.')
    parser.add_argument('--server-name',
                        type=str,
                        default='127.0.0.1',
                        help='Demo server name. Default: 127.0.0.1, which is only visible from the local computer.'
                        ' If you want other computers to access your server, use 0.0.0.0 instead.',
                        )

    args = parser.parse_args()
    return args


if __name__ == '__main__':
    args = _get_args()

    # Now you do not need to add "trust_remote_code=True"
    model = AutoModelForCausalLM.from_pretrained(
        args.checkpoint_path,
        torch_dtype="auto",
        device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(args.checkpoint_path)

    uvicorn.run(app, host=args.server_name, port=args.server_port, workers=1)

调用下接口：

[图片：接口调用结果]

可以看到接口已经返回内容了。

增加流式支持

这里需要安装 sse_starlette 的库，来支持流式的返回：

pip install sse_starlette

安装完把代码再改一下，通过参数 stream 来判断是否流式返回：

from contextlib import asynccontextmanager
from threading import Thread
import torch
import uvicorn
import time
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, BatchEncoding
from fastapi import FastAPI, HTTPException
from fastapi.middleware.cors import CORSMiddleware
from argparse import ArgumentParser
from typing import List, Literal, Optional, Union
from pydantic import BaseModel, Field
from sse_starlette.sse import EventSourceResponse


@asynccontextmanager
async def lifespan(app: FastAPI):  # collects GPU memory
    yield
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.ipc_collect()

app = FastAPI(lifespan=lifespan)

# 支持 CORS
app.add_middleware(
    CORSMiddleware,
    allow_origins=['*'],
    allow_credentials=True,
    allow_methods=['*'],
    allow_headers=['*'],
)


class ChatMessage(BaseModel):
    role: Literal['user', 'assistant', 'system']
    content: Optional[str]


class DeltaMessage(BaseModel):
    role: Optional[Literal['user', 'assistant', 'system']] = None
    content: Optional[str] = None


class ChatCompletionRequest(BaseModel):
    model: str
    messages: List[ChatMessage]
    stream: Optional[bool] = False


class ChatCompletionResponseChoice(BaseModel):
    index: int
    message: ChatMessage
    finish_reason: Literal['stop', 'length']


class ChatCompletionResponseStreamChoice(BaseModel):
    index: int
    delta: DeltaMessage
    finish_reason: Optional[Literal['stop', 'length']]


class ChatCompletionResponse(BaseModel):
    model: str
    object: Literal['chat.completion', 'chat.completion.chunk']
    choices: List[Union[ChatCompletionResponseChoice,
                        ChatCompletionResponseStreamChoice]]
    created: Optional[int] = Field(default_factory=lambda: int(time.time()))


@app.get("/")
async def index():

    return {"message": "Hello World"}


@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
async def create_chat_completion(request: ChatCompletionRequest):
    global model, tokenizer

    # 简单的错误校验
    if request.messages[-1].role != "user":
        raise HTTPException(status_code=400, detail="Invalid request")

    text = tokenizer.apply_chat_template(
        request.messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

    if request.stream:
        generate = predict(model_inputs, request.model)
        return EventSourceResponse(generate, media_type="text/event-stream")

    # Directly use generate() and tokenizer.decode() to get the output.
    # Use `max_new_tokens` to control the maximum output length.
    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]

    choice_data = ChatCompletionResponseChoice(
        index=0,
        message=ChatMessage(role="assistant", content=response),
        finish_reason="stop"
    )

    return ChatCompletionResponse(model=request.model, choices=[choice_data], object="chat.completion")


async def predict(model_inputs: BatchEncoding,  model_id: str):
    global model, tokenizer

    streamer = TextIteratorStreamer(
        tokenizer, skip_prompt=True, skip_special_tokens=True)
    generation_kwargs = dict(
        model_inputs, streamer=streamer, max_new_tokens=512)
    thread = Thread(target=model.generate, kwargs=generation_kwargs)

    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(role="assistant"),
        finish_reason=None
    )
    chunk = ChatCompletionResponse(model=model_id, choices=[
                                   choice_data], object="chat.completion.chunk")
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))

    thread.start()
    for new_text in streamer:
        choice_data = ChatCompletionResponseStreamChoice(
            index=0,
            delta=DeltaMessage(content=new_text),
            finish_reason=None
        )

        chunk = ChatCompletionResponse(model=model_id, choices=[
                                       choice_data], object="chat.completion.chunk")
        yield "{}".format(chunk.model_dump_json(exclude_unset=True))

    choice_data = ChatCompletionResponseStreamChoice(
        index=0,
        delta=DeltaMessage(),
        finish_reason="stop"
    )
    chunk = ChatCompletionResponse(model=model_id, choices=[
                                   choice_data], object="chat.completion.chunk")
    yield "{}".format(chunk.model_dump_json(exclude_unset=True))
    yield '[DONE]'


def _get_args():
    parser = ArgumentParser()

    parser.add_argument(
        '-c',
        '--checkpoint-path',
        type=str,
        default='Qwen1.5-0.5B-Chat',
        help='Checkpoint name or path, default to %(default)r',
    )

    parser.add_argument('--server-port',
                        type=int,
                        default=8000,
                        help='Demo server port.')
    parser.add_argument('--server-name',
                        type=str,
                        default='127.0.0.1',
                        help='Demo server name. Default: 127.0.0.1, which is only visible from the local computer.'
                        ' If you want other computers to access your server, use 0.0.0.0 instead.',
                        )

    args = parser.parse_args()
    return args


if __name__ == '__main__':
    args = _get_args()

    # Now you do not need to add "trust_remote_code=True"
    model = AutoModelForCausalLM.from_pretrained(
        args.checkpoint_path,
        torch_dtype="auto",
        device_map="auto"
    )
    tokenizer = AutoTokenizer.from_pretrained(args.checkpoint_path)

    uvicorn.run(app, host=args.server_name, port=args.server_port, workers=1)

再调用接口：

[图片：流式接口测试]

流式返回也支持了。

完善接口

接口更加完善的话可以参考官方文档或社区示例。

总结

至此，Qwen1.5 大模型已在 Windows 环境下完成本地部署及 API 服务化。通过上述步骤，您可以实现从环境搭建、模型加载到接口调用的全流程开发。建议在实际生产环境中进一步考虑安全性、并发处理及资源优化等问题。