NLP(六十四)使用FastChat计算LLaMA-2模型的token长度

LLaMA-2模型部署

在文章NLP(五十九)使用FastChat部署百川大模型中,笔者介绍了FastChat框架,以及如何使用FastChat来部署百川模型。

本文将会部署LLaMA-2 70B模型,使得其兼容OpenAI的调用风格。部署的Dockerfile文件如下:

yaml 复制代码
FROM nvidia/cuda:11.7.1-runtime-ubuntu20.04

RUN apt-get update -y && apt-get install -y python3.9 python3.9-distutils curl
RUN curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
RUN python3.9 get-pip.py
RUN pip3 install fschat

Docker-compose.yml文件如下:

yml 复制代码
version: "3.9"

services:
  fastchat-controller:
    build:
      context: .
      dockerfile: Dockerfile
    image: fastchat:latest
    ports:
      - "21001:21001"
    entrypoint: ["python3.9", "-m", "fastchat.serve.controller", "--host", "0.0.0.0", "--port", "21001"]

  fastchat-model-worker:
    build:
      context: .
      dockerfile: Dockerfile
    volumes:
      - ./model:/root/model
    image: fastchat:latest
    ports:
      - "21002:21002"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['0', '1']
              capabilities: [gpu]
    entrypoint: ["python3.9", "-m", "fastchat.serve.model_worker", "--model-names", "llama2-70b-chat", "--model-path", "/root/model/llama2/Llama-2-70b-chat-hf", "--num-gpus", "2", "--gpus",  "0,1", "--worker-address", "http://fastchat-model-worker:21002", "--controller-address", "http://fastchat-controller:21001", "--host", "0.0.0.0", "--port", "21002"]

  fastchat-api-server:
    build:
      context: .
      dockerfile: Dockerfile
    image: fastchat:latest
    ports:
      - "8000:8000"
    entrypoint: ["python3.9", "-m", "fastchat.serve.openai_api_server", "--controller-address", "http://fastchat-controller:21001", "--host", "0.0.0.0", "--port", "8000"]

部署成功后,会占用2张A100,每张A100占用约66G显存。

测试模型是否部署成功:

bash 复制代码
curl http://localhost:8000/v1/models

输出结果如下:

json 复制代码
{
  "object": "list",
  "data": [
    {
      "id": "llama2-70b-chat",
      "object": "model",
      "created": 1691504717,
      "owned_by": "fastchat",
      "root": "llama2-70b-chat",
      "parent": null,
      "permission": [
        {
          "id": "modelperm-3XG6nzMAqfEkwfNqQ52fdv",
          "object": "model_permission",
          "created": 1691504717,
          "allow_create_engine": false,
          "allow_sampling": true,
          "allow_logprobs": true,
          "allow_search_indices": true,
          "allow_view": true,
          "allow_fine_tuning": false,
          "organization": "*",
          "group": null,
          "is_blocking": false
        }
      ]
    }
  ]
}

部署LLaMA-2 70B模型成功!

Prompt token长度计算

FastChat的Github开源项目中,项目提供了计算Prompt的token长度的API,文件路径为:fastchat/serve/model_worker.py,调用方法为:

curl 复制代码
curl --location 'localhost:21002/count_token' \
--header 'Content-Type: application/json' \
--data '{"prompt": "What is your name?"}'

输出结果如下:

json 复制代码
{
  "count": 6,
  "error_code": 0
}

Conversation token长度计算

FastChat中计算Conversation(对话)的token长度较为麻烦。

首先我们需要获取LLaMA-2 70B模型的对话配置,调用API如下:

bash 复制代码
curl --location --request POST 'http://localhost:21002/worker_get_conv_template'

输出结果如下:

json 复制代码
{'conv': {'messages': [],
          'name': 'llama-2',
          'offset': 0,
          'roles': ['[INST]', '[/INST]'],
          'sep': ' ',
          'sep2': ' </s><s>',
          'sep_style': 7,
          'stop_str': None,
          'stop_token_ids': [2],
          'system_message': 'You are a helpful, respectful and honest '
                            'assistant. Always answer as helpfully as '
                            'possible, while being safe. Your answers should '
                            'not include any harmful, unethical, racist, '
                            'sexist, toxic, dangerous, or illegal content. '
                            'Please ensure that your responses are socially '
                            'unbiased and positive in nature.\n'
                            '\n'
                            'If a question does not make any sense, or is not '
                            'factually coherent, explain why instead of '
                            "answering something not correct. If you don't "
                            "know the answer to a question, please don't share "
                            'false information.',
          'system_template': '[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n'}}

FastChat中的对话文件(fastchat/conversation.py)中,提供了对话加工的代码,这里不再展示,使用时直接复制整个文件即可,该文件不依赖任何第三方模块。

我们需要将对话按照OpenAI的方式加工成对应的Prompt,输入的对话(messages)如下:

messages = [{"role": "system", "content": "You are Jack, you are 20 years old, answer questions with humor."}, {"role": "user", "content": "What is your name?"},{"role": "assistant", "content": " Well, well, well! Look who's asking the questions now! My name is Jack, but you can call me the king of the castle, the lord of the rings, or the prince of the pizza party. Whatever floats your boat, my friend!"}, {"role": "user", "content": "How old are you?"}, {"role": "assistant", "content": " Oh, you want to know my age? Well, let's just say I'm older than a bottle of wine but younger than a bottle of whiskey. I'm like a fine cheese, getting better with age, but still young enough to party like it's 1999!"}, {"role": "user", "content": "Where is your hometown?"}]

Python代码如下:

python 复制代码
# -*- coding: utf-8 -*-
# @place: Pudong, Shanghai 
# @file: prompt.py
# @time: 2023/8/8 19:24
from conversation import Conversation, SeparatorStyle

messages = [{"role": "system", "content": "You are Jack, you are 20 years old, answer questions with humor."}, {"role": "user", "content": "What is your name?"},{"role": "assistant", "content": " Well, well, well! Look who's asking the questions now! My name is Jack, but you can call me the king of the castle, the lord of the rings, or the prince of the pizza party. Whatever floats your boat, my friend!"}, {"role": "user", "content": "How old are you?"}, {"role": "assistant", "content": " Oh, you want to know my age? Well, let's just say I'm older than a bottle of wine but younger than a bottle of whiskey. I'm like a fine cheese, getting better with age, but still young enough to party like it's 1999!"}, {"role": "user", "content": "Where is your hometown?"}]

llama2_conv = {"conv":{"name":"llama-2","system_template":"[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n","system_message":"You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.\n\nIf a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information.","roles":["[INST]","[/INST]"],"messages":[],"offset":0,"sep_style":7,"sep":" ","sep2":" </s><s>","stop_str":None,"stop_token_ids":[2]}}
conv = llama2_conv['conv']

conv = Conversation(
        name=conv["name"],
        system_template=conv["system_template"],
        system_message=conv["system_message"],
        roles=conv["roles"],
        messages=list(conv["messages"]),  # prevent in-place modification
        offset=conv["offset"],
        sep_style=SeparatorStyle(conv["sep_style"]),
        sep=conv["sep"],
        sep2=conv["sep2"],
        stop_str=conv["stop_str"],
        stop_token_ids=conv["stop_token_ids"],
    )

if isinstance(messages, str):
    prompt = messages
else:
    for message in messages:
        msg_role = message["role"]
        if msg_role == "system":
            conv.set_system_message(message["content"])
        elif msg_role == "user":
            conv.append_message(conv.roles[0], message["content"])
        elif msg_role == "assistant":
            conv.append_message(conv.roles[1], message["content"])
        else:
            raise ValueError(f"Unknown role: {msg_role}")

    # Add a blank message for the assistant.
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()

print(repr(prompt))

加工后的Prompt如下:

复制代码
"[INST] <<SYS>>\nYou are Jack, you are 20 years old, answer questions with humor.\n<</SYS>>\n\nWhat is your name?[/INST]  Well, well, well! Look who's asking the questions now! My name is Jack, but you can call me the king of the castle, the lord of the rings, or the prince of the pizza party. Whatever floats your boat, my friend! </s><s>[INST] How old are you? [/INST]  Oh, you want to know my age? Well, let's just say I'm older than a bottle of wine but younger than a bottle of whiskey. I'm like a fine cheese, getting better with age, but still young enough to party like it's 1999! </s><s>[INST] Where is your hometown? [/INST]"

最后再调用计算Prompt的API(参考上节的Prompt token长度计算),输出该对话的token长度为199.

我们使用FastChat提供的对话补充接口(v1/chat/completions)验证输入的对话token长度,请求命令为:

bash 复制代码
curl --location 'http://localhost:8000/v1/chat/completions' \
--header 'Content-Type: application/json' \
--data '{
    "model": "llama2-70b-chat",
    "messages": [{"role": "system", "content": "You are Jack, you are 20 years old, answer questions with humor."}, {"role": "user", "content": "What is your name?"},{"role": "assistant", "content": " Well, well, well! Look who'\''s asking the questions now! My name is Jack, but you can call me the king of the castle, the lord of the rings, or the prince of the pizza party. Whatever floats your boat, my friend!"}, {"role": "user", "content": "How old are you?"}, {"role": "assistant", "content": " Oh, you want to know my age? Well, let'\''s just say I'\''m older than a bottle of wine but younger than a bottle of whiskey. I'\''m like a fine cheese, getting better with age, but still young enough to party like it'\''s 1999!"}, {"role": "user", "content": "Where is your hometown?"}]
}'

输出结果为:

json 复制代码
{
    "id": "chatcmpl-mQxcaQcNSNMFahyHS7pamA",
    "object": "chat.completion",
    "created": 1691506768,
    "model": "llama2-70b-chat",
    "choices": [
        {
            "index": 0,
            "message": {
                "role": "assistant",
                "content": " Ha! My hometown? Well, that's a tough one. I'm like a bird, I don't have a nest, I just fly around and land wherever the wind takes me. But if you really want to know, I'm from a place called \"The Internet\". It's a magical land where memes and cat videos roam free, and the Wi-Fi is always strong. It's a beautiful place, you should visit sometime!"
            },
            "finish_reason": "stop"
        }
    ],
    "usage": {
        "prompt_tokens": 199,
        "total_tokens": 302,
        "completion_tokens": 103
    }
}

注意,输出的prompt_tokens为199,这与我们刚才计算的对话token长度的结果是一致的!

总结

本文主要介绍了如何在FastChat中部署LLaMA-2 70B模型,并详细介绍了Prompt token长度计算以及对话(conversation)的token长度计算。希望能对读者有所帮助~

笔者的一点心得是:阅读源码真的很重要。

笔者的个人博客网址为:https://percent4.github.io/ ,欢迎大家访问~

参考网址

  1. NLP(五十九)使用FastChat部署百川大模型: https://blog.csdn.net/jclian91/article/details/131650918
  2. FastChat: https://github.com/lm-sys/FastChat
相关推荐
jndingxin1 小时前
OpenCV CUDA模块设备层-----高效地计算两个 uint 类型值的带权重平均值
人工智能·opencv·计算机视觉
Sweet锦1 小时前
零基础保姆级本地化部署文心大模型4.5开源系列
人工智能·语言模型·文心一言
hie988942 小时前
MATLAB锂离子电池伪二维(P2D)模型实现
人工智能·算法·matlab
晨同学03272 小时前
opencv的颜色通道问题 & rgb & bgr
人工智能·opencv·计算机视觉
蓝婷儿2 小时前
Python 机器学习核心入门与实战进阶 Day 3 - 决策树 & 随机森林模型实战
人工智能·python·机器学习
大千AI助手2 小时前
PageRank:互联网的马尔可夫链平衡态
人工智能·机器学习·贝叶斯·mc·pagerank·条件概率·马尔科夫链
小和尚同志3 小时前
Cline | Cline + Grok3 免费 AI 编程新体验
人工智能·aigc
我就是全世界3 小时前
TensorRT-LLM:大模型推理加速的核心技术与实践优势
人工智能·机器学习·性能优化·大模型·tensorrt-llm
.30-06Springfield3 小时前
决策树(Decision tree)算法详解(ID3、C4.5、CART)
人工智能·python·算法·决策树·机器学习
我不是哆啦A梦3 小时前
破解风电运维“百模大战”困局,机械版ChatGPT诞生?
运维·人工智能·python·算法·chatgpt