本地化部署deepseek，并搭建聊天窗口来时时chat

一、模型下载

首先到modelscope网站下载模型：魔搭社区

大概有十几个G，预计一个小时左右，耐心等待即可。下载完成后如下：

下载完成后编写代码进行调用

二、模型调用

调用流程很简单，先是指定模型路径（下载存放的位置），用transformers加载模型，然后提供提示词来调用模型，然后模型运行生成答案。代码结构如图：

其中，

Deepseek.py是用来调用模型

app.py是为了启动服务，供聊天窗口随时调用；

chat_client.py是开发聊天窗口来随时和deepseek进行对话

1、模型加载

直接撸代码：

import torch
import time
from transformers import AutoModelForCausalLM, AutoTokenizer

from utils.data_dir import root_dir

model_name = f"{root_dir}/DeepSeek-R1-Distill-Qwen-7B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

2、模型调用

直接撸代码：

def generate_response(prompt, max_new_tokens=1024):  # Format the prompt as a chat message
    messages = [{"role": "user", "content": prompt}]
    # Apply the chat template and tokenize the input
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    # Prepare the input for the model
    model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
    # 确保tokenizer中有正确的pad_token和eos_token
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
        tokenizer.pad_token_id = tokenizer.eos_token_id

    # 设置正确的pad_token_id和eos_token_id
    model.config.pad_token_id = tokenizer.pad_token_id
    model.config.eos_token_id = tokenizer.eos_token_id
    generated_ids = model.generate(
        **model_inputs,
        max_new_tokens=max_new_tokens,  # Control the length of the generated text
        pad_token_id=tokenizer.pad_token_id,  # 防止报填充和结束混淆错误
        eos_token_id=tokenizer.eos_token_id
    )
    # Decode the generated IDs to text
    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]
    return response


def chat(prompt):
    start_time = time.time()
    response = generate_response(prompt)
    end_time = time.time()
    print(f"Time taken for prompt：'{prompt}...{end_time - start_time: .2f}' seconds")
    print(f"answer：'{response}")
    return response

3、服务启动

直接撸代码：

#!/user/bin/env python3
#  -*- coding: utf-8 -*-
from pathlib import Path
from traceback import format_exc
from flask import jsonify, request, Flask

from deepseek import chat
from utils.logger import Log

os_file_name = Path(__file__).name

app = Flask(__name__)


@app.route("/chat/deepseek", methods=["POST"])
def chat_with_deepseek():
    try:
        # 获取参数
        request_params = request.get_json(force=True)
        question = request_params["question"]
        Log().info(os_file_name,  f"问题: {question}")
        # 反射获取方法
        try:
            # import module, func
            result = chat(question)
            think = result.split("</think>")[0]
            answer = result.split("</think>")[1]
        except (ModuleNotFoundError, AttributeError) as e:
            Log().info(os_file_name,  repr(e))
            return jsonify(result=1, resultNote=f"服务器繁忙……")
        # 返回执行结果
        return jsonify(result=0, resultNote="chat success", think=think, answer=answer)

    except Exception as e:
        Log().error(os_file_name,  repr(e))
        Log().error(os_file_name,  format_exc())
        return jsonify(result=2, resultNote="服务内部异常")


if __name__ == '__main__':
    app.run(host="0.0.0.0", debug=False, port=8080)

4、聊天窗口

直接撸代码：

#!/usr/bin/env python
# -*- coding:utf-8 -*-
# @Author: 人海中的海盗
# 在线聊天客户端
import tkinter
import tkinter.font as tkFont
import time
import requests
global clientSock


class ClientUI:
    # 初始化类的相关属性的构造函数
    def __init__(self):
        self.root = tkinter.Tk()
        self.root.title('Python客户端-deepseek咨询')
        # 窗口  4个frame面板布局
        self.frame = [tkinter.Frame(), tkinter.Frame(), tkinter.Frame(), tkinter.Frame()]
        # 滚动条
        self.chatTextScrollBar = tkinter.Scrollbar(self.frame[0])
        self.chatTextScrollBar.pack(side=tkinter.RIGHT, fill=tkinter.Y)
        # 显示Text，绑定滚动条
        ft = tkFont.Font(family='Fixdsys', size=11)
        self.chatText = tkinter.Listbox(self.frame[0], width=70, height=18, font=ft)
        self.chatText['yscrollcommand'] = self.chatTextScrollBar.set
        self.chatText.pack(expand=1, fill=tkinter.BOTH)
        self.chatTextScrollBar['command'] = self.chatText.yview()
        self.frame[0].pack(expand=1, fill=tkinter.BOTH)
        # 分开消息显示和消息输入
        label = tkinter.Label(self.frame[1], height=2)
        label.pack(fill=tkinter.BOTH)
        self.frame[1].pack(expand=1, fill=tkinter.BOTH)
        # 输入消息的滚动条
        self.inputTextScrollBar = tkinter.Scrollbar(self.frame[2])
        self.inputTextScrollBar.pack(side=tkinter.RIGHT, fill=tkinter.Y)

        ft = tkFont.Font(family='Fixdsys', size=11)
        self.inputText = tkinter.Text(self.frame[2], width=70, height=8, font=ft)
        self.inputText['yscrollcommand'] = self.inputTextScrollBar.set
        self.inputText.pack(expand=1, fill=tkinter.BOTH)
        self.inputTextScrollBar['command'] = self.chatText.yview()
        self.frame[2].pack(expand=1, fill=tkinter.BOTH)
        # 发送按钮
        self.sendButton = tkinter.Button(self.frame[3], text='发送', width=10, command=self.receive_message)
        self.sendButton.pack(expand=1, side=tkinter.BOTTOM and tkinter.RIGHT, padx=25, pady=5)
        # 关闭按钮
        self.closeButton = tkinter.Button(self.frame[3], text='关闭', width=10, command=self.close)
        self.closeButton.pack(expand=1, side=tkinter.RIGHT, padx=25, pady=5)
        self.frame[3].pack(expand=1, fill=tkinter.BOTH)

        self.buffer = 1024

    def close(self):
        self.root.destroy()

    # 接收消息
    def receive_message(self):
        message = self.inputText.get('1.0', tkinter.END)
        if message.strip() == "":
            self.chatText.insert(tkinter.END, '请输入问题后再发送……\n')
            return -1
        theTime = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
        self.chatText.insert(tkinter.END, '客观' + theTime + '问：\n' + message + '\n')
        json_q = {"question": message.strip()}
        try:
            response = requests.post(url="http://171.1.9.1:8080/chat/deepseek", json=json_q, timeout=300)
            result = response.json()
            if result["result"] == 0:
                self.flag = True
                self.server_msg = f'{result["answer"]} \n 所以经过我的思考，{result["answer"]}'
                the_time = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
                self.chatText.insert(tkinter.END, f'deep seek {the_time} 回答：{self.server_msg}\n')
                self.chatText.see(tkinter.END)
            else:
                self.flag = False
                self.chatText.insert(tkinter.END, '服务器繁忙\n')
        except Exception as e:
            self.chatText.insert(tkinter.END, f'服务器发生异常：{str(e)}\n')


if __name__ == '__main__':
    client = ClientUI()
    client.root.mainloop()

PS：

有个错误（pad_toekn_id to eos_token_id）要记录一下，如果踩坑的可以借鉴：

解决办法我已经放到代码里面了，就是在model.generate函数中指定pad_token_id=tokenizer.pad_token_id,  eos_token_id=tokenizer.eos_token_id。

最后我们看下运行效果（我只有一张4090显卡，7B的模型带不起来，调用一次起码一分钟）：