【AI实战】开源中文 llama2 来了,30 分钟搭建 130 亿参数大模型 Llama2-Chinese-13b-Chat

简介

环境配置

环境搭建

  • 系统环境

    • Ubuntu 20.04LTS
    • CUDA 11.8
    • Docker 18.09.5
  • 创建docker容器

拉取docker镜像

docker pull nvcr.io/nvidia/pytorch:21.08-py3

创建docker


nvidia-docker run -it -d \
    --name llama2_chinese \
    -p 15550:15550 \
    -p 15551:15551 \
    -v /xx/xx/llm:/notebooks \
    -e TZ='Asia/Shanghai' \
    --shm-size 16G \
    nvcr.io/nvidia/pytorch:21.08-py3
    
# docker exec -it llama2_chinese env LANG=C.UTF-8 /bin/bash

修改 /xx/xx/llm 为自己的路径

进入容器内:

docker exec -it llama2_chinese env LANG=C.UTF-8 /bin/bash
  • 安装conda

下载:

cd /notebooks
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh

安装:

bash Miniconda3-latest-Linux-x86_64.sh

安装提升安装即可

将miniconda加入PATH路径:

export PATH="/root/miniconda3/bin:$PATH"

创建 conda 环境:

conda create -n llama2 python=3.10.9
  • 安装依赖库
conda activate llama2
conda init

exit退出docker,重新进入docker

docker exec -it llama2_chinese  env LANG=C.UTF-8 /bin/bash
cd /notebooks
conda activate llama2

依赖安装

安装依赖库:

pip install -r requirements-2.txt  -i https://pypi.tuna.tsinghua.edu.cn/simple

代码及模型权重拉取

拉取 Llama2-Chinese

git clone https://github.com/FlagAlpha/Llama2-Chinese.git

github网站偶尔会抽风,需要耐心等待,如果失败了,执行 rm -rf Llama2-Chinese,再重新拉取

拉取 Llama2-Chinese-13b-Chat 模型权重及代码

cd Llama2-Chinese
git clone git clone https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat

由于权重文件特别大,如果失败了,执行 rm -rf Llama2-Chinese-13b-Chat,再重新拉取。
建议中午时间拉取,速度比较快,大概 1 小时(和你的网络带宽有很大关系!)。

文件大小查看:

du -sh Llama2-Chinese-13b-Chat

输出:

25G    Llama2-Chinese-13b-Chat

如果文件大小不一致,或者太小,说明权重文件下载识别,
请执行: rm -rf Llama2-Chinese-13b-Chat, 再重新拉取
或者 按照后面的方法去下载权重文件

查看文件列表:

ls -l Llama2-Chinese-13b-Chat/

输出:

# ls -l Llama2-Chinese-13b-Chat/
total 25421840
-rw-r--r-- 1 root root       1514 Jul 27 06:33 README.md
-rw-r--r-- 1 root root        683 Jul 27 06:33 config.json
-rw-r--r-- 1 root root        327 Jul 27 07:12 down.sh
-rw-r--r-- 1 root root        175 Jul 27 06:33 generation_config.json
-rw-r--r-- 1 root root 9948728430 Jul 24 23:12 pytorch_model-00001-of-00003.bin
-rw-r--r-- 1 root root 9904165024 Jul 24 12:44 pytorch_model-00002-of-00003.bin
-rw-r--r-- 1 root root 6178983625 Jul 24 12:14 pytorch_model-00003-of-00003.bin
-rw-r--r-- 1 root root      33444 Jul 27 06:33 pytorch_model.bin.index.json
-rw-r--r-- 1 root root        414 Jul 27 06:33 special_tokens_map.json
-rw-r--r-- 1 root root        131 Jul 27 06:33 tokenizer.model
-rw-r--r-- 1 root root        749 Jul 27 06:33 tokenizer_config.json

【】如果大文件下载识别,按照下面的方法下载模型文件;

wget https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat/resolve/main/pytorch_model-00001-of-00003.bin
wget https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat/resolve/main/pytorch_model-00002-of-00003.bin
wget https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat/resolve/main/pytorch_model-00003-of-00003.bin

如果还是没法下载,参考下面的“国内 Llama2 最新下载地址” 进行下载

终端测试

进入python环境:

python3

输入代码:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained('Llama2-Chinese-13b-Chat',device_map='auto',torch_dtype=torch.float16,load_in_8bit=True)
model =model.eval()
tokenizer = AutoTokenizer.from_pretrained('Llama2-Chinese-13b-Chat',use_fast=False)
tokenizer.pad_token = tokenizer.eos_token
input_ids = tokenizer(['<s>Human: 介绍一下深圳\n</s><s>Assistant: '], return_tensors="pt",add_special_tokens=False).input_ids.to('cuda')        
generate_input = {
    "input_ids":input_ids,
    "max_new_tokens":512,
    "do_sample":True,
    "top_k":50,
    "top_p":0.95,
    "temperature":0.3,
    "repetition_penalty":1.3,
    "eos_token_id":tokenizer.eos_token_id,
    "bos_token_id":tokenizer.bos_token_id,
    "pad_token_id":tokenizer.pad_token_id
}
generate_ids  = model.generate(**generate_input)
text = tokenizer.decode(generate_ids[0])
print(text)

页面测试

使用 gradio 搭建页面

安装 gradio

pip install gradio -i https://pypi.tuna.tsinghua.edu.cn/simple

加载模型并启动服务

修改端口;

vi /notebooks/Llama2-Chinese/examples/chat_gradio.py

到94行:

    demo.queue().launch(share=False, debug=True, server_name="0.0.0.0")

修改为:

    demo.queue().launch(share=False, debug=True, server_name="0.0.0.0", server_port=15550)

启动脚本:

python examples/chat_gradio.py --model_name_or_path Llama2-Chinese-13b-Chat

如果出现下面的错误:

  File "/notebooks/Llama2-Chinese/examples/chat_gradio.py", line 94
    demo.queue().launch(share=False, debug=True, server_name="0.0.0.0")
                                               ^
SyntaxError: invalid character ',' (U+FF0C)

则按照下面的步骤修改代码:

vi /notebooks/Llama2-Chinese/examples/chat_gradio.py
:94 
修改中文逗号,为英文逗号,

94    demo.queue().launch(share=False, debug=True, server_name="0.0.0.0")
=>
94    demo.queue().launch(share=False, debug=True, server_name="0.0.0.0")
  • 测试

    浏览器打开地址:http://10.192.x.x:15550/

国内 Llama2 最新下载地址

Llama2-7B官网版本:https://pan.xunlei.com/s/VN_kR2fwuJdG1F3CoF33rwpIA1?pwd=z9kf

Llama2-7B-Chat官网版本:https://pan.xunlei.com/s/VN_kQa1_HBvV-X9QVI6jV2kOA1?pwd=xmra

Llama2-13B官网版本:https://pan.xunlei.com/s/VN_izibaMDoptluWodzJw4cRA1?pwd=2qqb

Llama2-13B-Chat官网版本:https://pan.xunlei.com/s/VN_iyyponyapjIDLXJCNfqy7A1?pwd=t3xw

Llama2-7B Hugging Face版本:https://pan.xunlei.com/s/VN_t0dUikZqOwt-5DZWHuMvqA1?pwd=66ep

Llama2-7B-Chat Hugging Face版本:https://pan.xunlei.com/s/VN_oaV4BpKFgKLto4KgOhBcaA1?pwd=ufir

Llama2-13B Hugging Face版本:https://pan.xunlei.com/s/VN_yT_9G8xNOz0SDWQ7Mb_GZA1?pwd=yvgf

Llama2-13B-Chat Hugging Face版本:https://pan.xunlei.com/s/VN_yA-9G34NGL9B79b3OQZZGA1?pwd=xqrg

参考

1.https://github.com/FlagAlpha/Llama2-Chinese
2.https://huggingface.co/FlagAlpha/Llama2-Chinese-13b-Chat/tree/main
3.https://github.com/facebookresearch/llama
4.https://huggingface.co/meta-llama

猜你喜欢

转载自blog.csdn.net/zengNLP/article/details/131965453