tokenizer = AutoTokenizer.from_pretrained("model/chatglm-6b-int4", trust_remote_code=True) 我已经在目录下放置了model/chatglm-6b-int4文件夹，但模型坚持从.cache中导入，并报错： ModuleNotFoundError: No module named 'transformers_modules.model/chatglm-6b-int4'
Environment- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA Support (`python -c "import torch; print(torch.cuda.is_available())"`) :我的也是从缓存导入的，你是怎么让 chatglm-6b-int4 在 GPU 上跑的？我这边只能 CPU 推断
升级transfromers到更新的版本，我这边有效
@2019switch 升级下确实可以跑起来了，但为什么感觉4bit量化后显存占用少了但推断更慢了呢
@rxy1212 你好，先请教一下你是怎么下载预训练权重的额？好像huggingface登不了额，另外ｉnt4需要下载哪些文件呢？谢谢
@Albert337
import os
from huggingface_hub import login, logout, hf_hub_download
repo_id = 'THUDM/chatglm-6b-int4'
local_dir = './chatglm-6b-int4'
token = os.getenv('HUGGINGFACE_TOKEN')
login(token=token, add_to_git_credential=True)
hf_hub_download(repo_id=repo_id, filename="ice_text.model", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="configuration_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="config.json", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="modeling_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="quantization.py", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="quantization_kernels.c", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="quantization_kernels_parallel.c", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="tokenization_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False)
hf_hub_download(repo_id=repo_id, filename="tokenizer_config.json", local_dir=local_dir, local_dir_use_symlinks=False)
logout()可以用脚本下载，模型在 huggingface 上都有 https://huggingface.co/THUDM/chatglm-6b-int4
@Albert337
import os from huggingface_hub import login, logout, hf_hub_download repo_id = 'THUDM/chatglm-6b-int4' local_dir = './chatglm-6b-int4' token = os.getenv('HUGGINGFACE_TOKEN') login(token=token, add_to_git_credential=True) hf_hub_download(repo_id=repo_id, filename="ice_text.model", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="pytorch_model.bin", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="configuration_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="config.json", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="modeling_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="quantization.py", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="quantization_kernels.c", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="quantization_kernels_parallel.c", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="tokenization_chatglm.py", local_dir=local_dir, local_dir_use_symlinks=False) hf_hub_download(repo_id=repo_id, filename="tokenizer_config.json", local_dir=local_dir, local_dir_use_symlinks=False) logout()可以用脚本下载，模型在 huggingface 上都有 https://huggingface.co/THUDM/chatglm-6b-int4
首先很感谢你的回答，但是我这边还是显示requests.exceptions.ConnectionError，用了科学上网好像还是不行
@Albert337 这个是这样的，你多试几次就好了(报错就一直重复运行脚本)，一般前面两个大文件下完后面的文件又要下一次
@rxy1212 ,确实也试了很多次，还是没解决。不知道有啥稳定的下载办法没？或者有直接下载好放在云盘链接的没
升级transfromers到更新的版本，我这边有效
我这边试了不行
transformers已经升级到了4.31.0
一使用int8版本模型就会跑到.cache目录下面去，但使用fp16版本模型却自始至终没有这个问题