Langchain-Chatchat基于WSL安装记录

请先根据官方Readme操作

https://github.com/chatchat-space/Langchain-Chatchat

1. 安装Windows显卡驱动就绪，CUDA 11.8，环境变量PATH配置好
2. 安装WSL2 Ubuntu
3. 安装Miniconda
4. MiniConda创建venv环境，python=3.11
5. 手动根据requirements.txt到阿里镜像安装对应cuda的torch、torchaudio、torchvision、xformers版本
6. 根据官方文档走

• 报错1

2024-03-07 22:46:16,792 - SentenceTransformer.py[line:66] - INFO: Load pretrained SentenceTransformer: /mnt/d/ai/models/bge-large-zh 2024-03-07 22:46:22,766 - embeddings_api.py[line:39] - ERROR: Error while deserializing header: HeaderTooLarge AttributeError: ‘NoneType’ object has no attribute ‘conjugate’

解决方法：bge-large-zh模型需要把model.safetensors下载下来，包括ChatGLM3-6B也是，默认没有读取pytorch_model-XXX.bin，而是读取.safetensors

• 报错2

  2024-03-07 23:08:47,000 - utils.py[line:162] - ERROR: ImportError: 为文件/mnt/d/ai/Langchain-Chatchat/knowledge_base/samples/content/llm/img/大模型应用技术原理-幕布图片-580318-260070.jpg查找加载器RapidOCRLoader时出错：libGL.so.1: cannot open shared object file: No such file or directory

解决方法：pip install opencv-python-headless

• 报错3

  Could not locate the tokenization_chatglm.py inside THUDM/chatglm3-6b.

解决方法：https://github.com/THUDM/ChatGLM3/discussions/894

• 报错4

(‘samples’, ’test_files/langchain-ChatGLM_open.jsonl’, ‘从文件 samples/test_files/langchain-ChatGLM_open.jsonl 加载文档时出错：jq package not found, please install it with pip install jq’)

解决方法：pip install jq