Debug Records

Records for errors or bugs met before

CondaVerificationError: The package for pytorch located at...
- reason: cache of installation packages not cleaned
- solution: conda clean --all
- refer: https://blog.csdn.net/weixin_42069606/article/details/111052017
[pytorch]RuntimeError: Distributed package doesn't have NCCL built in
- reason: using torch.nn.distributed which requires NCCL not supporting Win
- solution: download NCCL(https://developer.nvidia.com/nccl/nccl-download) in wsl
- refer: installation guide: https://docs.nvidia.com/deeplearning/nccl/install-guide/index.html
[pytorch]torch.distributed.DistNetworkError: Unknown error
- reason: exclusive problem for torch2.2.1 in Win
- solution: upgrade torch to ~>=2.2.2
- refer: https://github.com/pytorch/pytorch/issues/118868, could be derived from https://github.com/pytorch/pytorch/issues/118737#issuecomment-1920785766
TypeError: add_code_sample_docstrings() got an unexpected keyword argument 'tokenizer_class'
- reason: deprecated api
- solution: change tokenizer_class to processor_class
- refer: https://github.com/openai/gpt-2/issues/299
Using bert-base-uncased throws ValueError: Connection error, and we cannot find the requested files in the cached path. Please try again or make sure your Internet connection is on.
- reason: fail to download bert-base-uncased in China Mainland
- solution: manual download and specify the model directory
- refer: https://blog.csdn.net/GCTTTTTT/article/details/136055023