当前位置:   article > 正文

用通俗易懂的方式讲解大模型:ChatGLM3-6B 部署指南_chatglm3-6b搭建

chatglm3-6b搭建

最近智谱 AI 对底层大模型又进行了一次升级,ChatGLM3-6B 正式发布,不仅在性能测试和各种测评的数据上有显著提升,还新增了一些新功能,包括工具调用、代码解释器等,最重要的一点是还是保持 6B 的这种低参数量,让我们可以在消费级的显卡上部署大语言模型(LLM)。

本文将对 ChatGLM3-6B 的部署做一次详细介绍,让更多人可以体验这个 LLM 的有趣功能。

环境安装

首先下载 ChatGLM3 的代码仓库,并安装相关的依赖。

git clone https://github.com/THUDM/ChatGLM3
cd ChatGLM3
pip install -r requirements.txt
  • 1
  • 2
  • 3

然后下载 ChatGLM3-6B 的模型文件,以下是笔者常用的 HuggingFace 下载方式。

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/THUDM/chatglm3-6b
cd chatglm3-6b
wget "https://huggingface.co/THUDM/chatglm3-6b/resolve/main/pytorch_model-00001-of-00007.bin"
wget "https://huggingface.co/THUDM/chatglm3-6b/resolve/main/pytorch_model-00002-of-00007.bin"
...
  • 1
  • 2
  • 3
  • 4
  • 5

ChatGLM3-6B 的模型文件在ModelScope[1]上也有提供下载,如果 HuggingFace 无法访问的话,可以从这上面下载。

git lfs install
git clone https://www.modelscope.cn/ZhipuAI/chatglm3-6b.git
  • 1
  • 2

通俗易懂讲解大模型系列

技术交流

建了大模型技术交流群! 想要学习、技术交流、获取如下原版资料的同学,可以直接加微信号:mlc2060。加的时候备注一下:研究方向 +学校/公司+CSDN,即可。然后就可以拉你进群了。

方式①、微信搜索公众号:机器学习社区,后台回复:加群
方式②、添加微信号:mlc2060,备注:来自CSDN + 技术交流

在这里插入图片描述

部署 WebUI 服务

ChatGLM3-6B 提供了两种 WebUI,分别是 Gradio 和 Streamlit。

Graido 服务

在启动 Gradio 服务之前需要先修改web_demo.py文件,将里面的模型地址改成本地的地址:

-tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm3-6b", trust_remote_code=True)
-model = AutoModel.from_pretrained("THUDM/chatglm3-6b", trust_remote_code=True).cuda()
+tokenizer = AutoTokenizer.from_pretrained("/root/autodl-tmp/chatglm3-6b", trust_remote_code=True)
+model = AutoModel.from_pretrained("/root/autodl-tmp/chatglm3-6b", trust_remote_code=True).cuda()
  • 1
  • 2
  • 3
  • 4

然后执行以下命令启动 Gradio 服务,服务启动后在浏览器中可以访问该服务进行对话:

python web_demo.py
  • 1

如果是使用 AutoDL 进行部署的话,可以将服务的端口设置6006(AutoDL 的开放端口),然后通过 AutoDL 的自定义服务进行访问。

图片

图片

Streamlit 服务

在启动 Streamlit 服务之前需要修改web_demo2.py,将里面的模型地址改成本地的地址:

-model_path = "THUDM/chatglm3-6b"
+model_path = "/root/autodl-tmp/chatglm3-6b"
  • 1
  • 2

然后执行以下命令启动 Streamlit 服务,服务启动后在浏览器中可以访问该服务:

streamlit run web_demo2.py
  • 1

如果是使用 AutoDL 进行部署的话,即使使用其开放端口(6006)也无法正常访问 Streamlit 的页面,页面会一直停留在Please wait...的提示中。因为 Streamlit 没有像 Gradio 那种内网代理功能,Gradio 在启动服务时可以通过 share=True 参数来生成一个公网链接,这个链接会代理到服务器的内部服务,这样在外部也可以正常访问 Gradio 服务。而 Streamlit 没有这种功能,所以我们需要通过 Ngrok 这个工具来实现内网穿透,将内网的服务代理到公网上,这样就可以正常访问页面了。

Ngrok 安装使用(可选)

首先在 Ngrok[2] 官网上查看安装命令,我们以 Linux 系统为例,有多种方式可以安装,包括压缩包下载、APT 安装、Snap 安装,这里我们使用 APT 安装,执行以下命令:

curl -s https://ngrok-agent.s3.amazonaws.com/ngrok.asc | sudo tee /etc/apt/trusted.gpg.d/ngrok.asc >/dev/null && echo "deb https://ngrok-agent.s3.amazonaws.com buster main" | sudo tee /etc/apt/sources.list.d/ngrok.list && sudo apt update && sudo apt install ngrok
  • 1

Ngrok 安装完成后,需要到它的官网上注册一个账号,然后在Your Authtoken菜单中获取 Authtoken,这个 Authtoken 用于验证用户身份,可以通过以下命令将 Authtoken 设置到本地。

ngrok config add-authtoken your-ngrok-authtoken # 这里替换成你的 Authtoken
  • 1

然后执行以下命令,通过 Ngrok 代理本地的 Streamlit 服务。

$ ngrok http 8501 # streamlit 默认端口为 8501

## 服务窗口
ngrok                                                                                                                   (Ctrl+C to quit)

Introducing Always-On Global Server Load Balancer: https://ngrok.com/r/gslb

Session Status                online
Account                       zhaozhiming (Plan: Free)
Version                       3.3.5
Region                        Japan (jp)
Latency                       -
Web Interface                 http://127.0.0.1:4040
Forwarding                    https://e7d9-36-111-143-226.ngrok-free.app -> http://localhost:8501

Connections                   ttl     opn     rt1     rt5     p50     p90
                              0       0       0.00    0.00    0.00    0.00
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11
  • 12
  • 13
  • 14
  • 15
  • 16
  • 17

最后我们通过窗口中的https://e7d9-36-111-143-226.ngrok-free.app地址就可以访问 Streamlit 服务了。

图片

部署 API 服务

启动 API 服务,服务的默认端口是 7861:

python openai_api.py
  • 1

该服务是兼容 OpenAI 接口,可以通过调用 OpenAI API 的方式来调用接口,注意要传递model参数,值为gpt-3.5-turbo

curl -X 'POST' \
  'https://localhost:7861/v1/chat/completions' \
  -H 'accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
  "model": "gpt-3.5-turbo",
  "messages": [
    {
      "role": "user",
      "content": "你好"
  ]
}'

# 返回结果
{
  "model": "gpt-3.5-turbo",
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "\n 你好!我是人工智能助手 ChatGLM3-6B,很高兴见到你,欢迎问我任何问题。",
        "metadata": null,
        "tools": null
      },
      "finish_reason": "stop",
      "history": null
    }
  ],
  "created": 1698825945,
  "usage": {
    "prompt_tokens": 8,
    "total_tokens": 38,
    "completion_tokens": 30
  }
}
  • 1
  • 2
  • 3
  • 4
  • 5
  • 6
  • 7
  • 8
  • 9
  • 10
  • 11
  • 12
  • 13
  • 14
  • 15
  • 16
  • 17
  • 18
  • 19
  • 20
  • 21
  • 22
  • 23
  • 24
  • 25
  • 26
  • 27
  • 28
  • 29
  • 30
  • 31
  • 32
  • 33
  • 34
  • 35
  • 36
  • 37

部署综合 Demo 服务

ChatGLM3-6B 还提供了一个综合的 Demo,包含了对话、工具调用、代码解释器等功能,我们来部署这个 Demo 服务。

首先可以按照官方文档说明新建一个 python 环境,然后安装相关依赖:

cd composite_demo
conda create -n chatglm3-demo python=3.10
conda activate chatglm3-demo
pip install -r requirements.txt
  • 1
  • 2
  • 3
  • 4

安装 Jupyter 内核和设置本地模型路径,然后启动 WebUI 服务:

ipython kernel install --name chatglm3-demo --user
export MODEL_PATH=/root/autodl-tmp/chatglm3-6b
streamlit run main.py
  • 1
  • 2
  • 3

Demo 里面除了对话功能外,还有工具调用和代码解释器功能,初始工具有两个,一个是天气查询,还有一个是随机数生成:

图片

图片

需要注意一点是:ChatGLM3-6B-32K 模型是没有工具调用功能的,只有 ChatGLM-6B 模型才有。

下面是代码解释器功能的截图,可以画爱心,还可以画饼图:

图片

图片

总结

上面介绍的是官方的部署,其实使用 FastChat 来部署更加简单,这种方式可以参考我之前的这篇文章,但是用 FastChat 可能无法使用 ChatGLM3-6B 的工具调用和代码解释器的功能。希望这篇文章能够帮助到大家,如果有什么问题可以在评论区留言。

关注我,一起学习各种人工智能和 AIGC 新技术,欢迎交流,如果你有什么想问想说的,欢迎在评论区留言。

参考:

[1] ModelScope: https://modelscope.cn

[2] Ngrok: https://ngrok.com/

声明:本文内容由网友自发贡献,不代表【wpsshop博客】立场,版权归原作者所有,本站不承担相应法律责任。如您发现有侵权的内容,请联系我们。转载请注明出处:https://www.wpsshop.cn/w/花生_TL007/article/detail/137314
推荐阅读
  

闽ICP备14008679号