从HuggingFace缓存加载Wizard-Vicuna-30B-Uncensored模型失败求助

1. 已确认模型存在本地缓存 #

执行huggingface-cli scan-cache命令，输出显示模型已下载：

REPO ID                                REPO TYPE SIZE ON DISK NB FILES LAST_ACCESSED LAST_MODIFIED REFS LOCAL PATH                                                                           
-------------------------------------- --------- ------------ -------- ------------- ------------- ---- ------------------------------------------------------------------------------------ 
ehartford/Wizard-Vicuna-30B-Uncensored model           130.1G       16 23 hours ago  3 weeks ago   main /home/hanyuan/.cache/huggingface/hub/models--ehartford--Wizard-Vicuna-30B-Uncensored 
openai/clip-vit-base-patch32           model           608.9M        8 10 hours ago  2 weeks ago   main /home/hanyuan/.cache/huggingface/hub/models--openai--clip-vit-base-patch32           

Done in 0.0s. Scanned 2 repo(s) for a total of 130.7G.

2. 加载模型的代码 #

model_name_or_path = "ehartford/Wizard-Vicuna-30B-Uncensored"
# model_name_or_path = "mistralai/Mistral-7B-Instruct-v0.1"
cache_path = "~/.cache/huggingface/hub/models--ehartford--Wizard-Vicuna-30B-Uncensored"

model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, cache_dir=cache_path)

3. 报错信息 #

FileNotFoundError: [Errno 2] No such file or directory: '../../blobs/2122720188751bb5b1d1a28480cbca2d216c63f2' -> '/home/hanyuan/.cache/huggingface/hub/models--ehartford--Wizard-Vicuna-30B-Uncensored/models--ehartford--Wizard-Vicuna-30B-Uncensored/snapshots/24768a6df2e4694b960ea84d2f304ad1ae46804d/config.json'

4. 排查情况 #

通过tree .查看缓存目录结构，确认config.json文件存在，但加载时无法读取。

解决方案 #

• 核心问题：cache_dir参数需要指向HuggingFace的根缓存目录（~/.cache/huggingface/hub），而非单个模型的具体目录。原代码中指定单个模型路径后，库会试图在这个路径下再创建一层模型目录，导致路径嵌套错误，无法找到文件。

• 修正后的代码（方法一）：指定根缓存目录

model_name_or_path = "ehartford/Wizard-Vicuna-30B-Uncensored"
# 指向HuggingFace根缓存目录
cache_path = "~/.cache/huggingface/hub"

model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16, cache_dir=cache_path)

• 修正后的代码（方法二）：不指定cache_dir，使用默认路径
  HuggingFace会自动读取~/.cache/huggingface/hub作为默认缓存目录，直接省略参数即可：

model_name_or_path = "ehartford/Wizard-Vicuna-30B-Uncensored"

model = AutoModelForCausalLM.from_pretrained(model_name_or_path, torch_dtype=torch.float16)

• 额外排查步骤：
  • 检查缓存目录的权限，确保当前用户拥有读写权限
  • 若问题持续，可执行huggingface-cli delete-cache清理损坏缓存后重新下载模型（注意提前备份重要文件）

内容的提问来源于stack exchange，提问作者HankY