LLAMA2（Meta大语言模型）可运行整合包的下载与安装

Windows10+消费级显卡可用

一、工程

github工程：

https://github.com/facebookresearch/llama

LLAMA2可运行整合包（Windows10+消费级显卡可用）

*现在只包括7B模型，会补充其他模型

文件夹下载

百度网盘链接：https://pan.baidu.com/s/1g1xknRdeHWPsAeZl9urXxg?pwd=ig7p

压缩包下载

百度网盘链接：https://pan.baidu.com/s/19gPNTTPc6cAdM2DLfmaE0w?pwd=w2me

夸克网盘链接：https://pan.quark.cn/s/b6ceb2c6211b

二、安装步骤

1，环境

在Linux机器上安装机器学习的环境

2，下载

运行代码下载工程

git clone https://github.com/facebookresearch/llama

3，下载模型

1）进入工程目录运行download.sh

bash download.sh

2）填入收到的邮件中的链接

https://download.llamameta.net/*?Policy=ey…

3）选择要下载的模型

6，安装

在目录\llama2 下，执行

         pip install -e .

三、运行

1，参数：

根据模型选择MP值

2、根据操作系统修改文件

1）在目录..\llama\generation.py 下

根据你的系统，把generation_linux.py或者generation_win.py修改为generation.py

或者在generation.py中直接修改，在62行：

if not torch.distributed.is_initialized():

   #torch.distributed.init_process_group("nccl") # linux

   torch.distributed.init_process_group("gloo")

2）如果不修改，可能出现错误

2，执行：

1）执行命令行

运行程序

torchrun --nproc_per_node 1 example_text_completion_fortest.py --ckpt_dir llama-2-7b  --tokenizer_path tokenizer.model --max_seq_len 128 --max_batch_size 4

2）解释：

关于命令行的解释：

3）执行结果

> initializing model parallel with size 1

> initializing ddp with size 1

> initializing pipeline with size 1

Loaded in 139.44 seconds

I have a website about learning AI technology name cyberxz.com, including knowledge about AI and machine learning.

> 1. AI, Machine Learning, Deep Learning, Natural Language Processing (NLP), etc. 2. AI applications in the real world, such as AI in self-driving cars, AI in healthcare, AI in finance, etc. 3. AI and

==================================

A brief message introducing the website:

        Hi everyone, this is a website

        ## 1. Introduction to the website

        This is a website with a brief introduction to the website.

        ## 2. Description of the website

        This is a website with a description of the website.

        ## 3. Detailed explanation of the website

==================================

4）修改prompt

在文件中修改

• example_text_completion.py
• example_chat_completion.py
• example_text_completion_fortest.py（我自己添加测试的）

prompts = [

   # For these prompts, the expected answer is the natural continuation of the prompt

   "I have a website about learning AI technology name cyberxz.com, including knowledge about AI and machine learning. ",

   """A brief message introducing the website:

   Hi everyone, this is a website

   """,

]