LLaMa2开源基模型申请、下载
1 申请模型
注意地区限制，选择HK或其他国家；组织可以填no organization。
2 下载模型
提交申请后一般几分钟就会通过，邮件或者meta官网会显示下载指导。
访问llama github项目
git clone github.com/facebookres…
设置+x权限
chmod +x download.sh
执行download.sh，按照提示操作，填入下载指导中的URL
(liyan_gptq_p38) root@ubuntu:/home/liyan/llm_datas/models/llama2-7b/llama# ls
CODE_OF_CONDUCT.md download.sh example_text_completion.py llama README.md Responsible-Use-Guide.pdf UPDATES.md
CONTRIBUTING.md example_chat_completion.py LICENSE MODEL_CARD.md requirements.txt setup.py USE_POLICY.md
(liyan_gptq_p38) root@ubuntu:/home/liyan/llm_datas/models/llama2-7b/llama# ./download.sh
Enter the URL from email: https://download.llamameta.net/*?XXXXXX
Enter the list of models to download without spaces (7B,13B,70B,7B-chat,13B-chat,70B-chat), or press Enter for all: 7B
Downloading LICENSE and Acceptable Usage Policy
--2024-05-30 21:39:58-- https://download.llamameta.net/LICENSE?Policy=XXXXXX
Resolving proxy.huawei.com (proxy.huawei.com)... 172.19.90.131
Connecting to proxy.huawei.com (proxy.huawei.com)|172.19.90.131|:8080... connected.
ERROR: cannot verify download.llamameta.net's certificate, issued by ‘CN=Huawei Web Secure Internet Gateway CA,OU=IT,O=Huawei,L=Shenzhen,ST=GuangDong,C=cn’:
Self-signed certificate encountered.
To connect to download.llamameta.net insecurely, use `--no-check-certificate'.
根据报错提示，修改download.sh脚本，添加--no-check-certificate参数，重新执行。
快速下载huggingface模型
step1: 安装依赖
pip install -U huggingface_hub
pip install -U hf-transfer
注意：huggingface_hub依赖于 Python>=3.8，hf_transfer 依附并兼容 huggingface-cli。
step2：设置ENV
export HF_ENDPOINT=https://hf-mirror.com
# 开启
export HF_HUB_ENABLE_HF_TRANSFER=1
# 关闭
export HF_HUB_ENABLE_HF_TRANSFER=0
step3: 基本用法
huggingface-cli命令行方式
# 获取相关信息
huggingface-cli download --help
# 下载单个文件（repo_id + 文件名）
huggingface-cli download gpt2 config.json
# 下载整个存储库（repo_id）
huggingface-cli download HuggingFaceH4/zephyr-7b-beta
# 下载多个文件
huggingface-cli download gpt2 config.json model.safetensors
# 下载dataset/space
huggingface-cli download HuggingFaceH4/ultrachat_200k --repo-type dataset
huggingface-cli download HuggingFaceH4/zephyr-chat --repo-type space
# 下载存储库里单个文件夹下的单个文件
huggingface-cli download stabilityai/stable-diffusion-xl-base-1.0 tokenizer/vocab.json
# 常用命令，保存到本地，--local-dir-use-symlinks 可以设置为True
huggingface-cli download --resume-download --local-dir-use-symlinks False stabilityai/stable-diffusion-xl-base-1.0 tokenizer/vocab.json --local-dir models
下载脚本方式
可以配置内网代理，关闭SSL校验
# coding: utf-8
import os
import urllib3
urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
import requests
from huggingface_hub import configure_http_backend, snapshot_download
os.environ["HF_ENDPOINT"] = "https://hf-mirror.com" # 镜像站加速
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
def backend_factory() -> requests.Session:
session = requests.Session()
proxies = {"http": "https://工号:密码@proxyhk.xxx.com:8080",
"https": "http://工号:密码@proxyhk.xxxx.com:8080"}
session.verify = False
session.trust_env = False
session.proxies.update(proxies)
return session
# Set it as the default session factory
configure_http_backend(backend_factory=backend_factory)
snapshot_download(
repo_type=None, # 要下载的类型是模型还是数据集，可选[None, 'model', 'dataset', 'space']
repo_id="openbmb/MiniCPM-2B-sft-bf16", # 仓库名
revision="main", # 分支名
local_dir="./model/MiniCPM-2B-sft-bf16", # 本地存储地址
token="xxx" # huggingface登录获取
)
实际使用时，发现还会报错[SSL: CERTIFICATE_VERIFY_FAILED]
下载huggingface报错[SSL: CERTIFICATE_VERIFY_FAILED]
执行下载时报错：ssl.SSLCertVerificationError: [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: self signed certificate in certificate chain (_ssl.c:1131) 解决措施：关掉证书校验
/lib/python3.11/site-packages/requests/adapters.py, 在adapters.py中找到 send方法，修改默认verify=False。
< def send(self, request, stream=False, timeout=None, verify=True,
---
> def send(self, request, stream=False, timeout=None, verify=False,
同理对sessions.py request()方法，修改默认verify=False。
< hooks=None, stream=None, verify=None, cert=None, json=None):
---
> hooks=None, stream=None, verify=False, cert=None, json=None):
