更多信息详见我的另一篇文档:
介绍
加载权重错误
从本地加载 stable-diffusion 时,
model_id = 'D:\code\git_diffusion_repo\sd\stable-diffusion-v1-5'
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16, revision="fp16")
出现 加载文件错误:
Loading pipeline components...: 0%| | 0/7 [00:00<?, ?it/s]
Traceback (most recent call last):File "D:\code\Github_code\sd\stable-diffusion-main\加载权重.py", line 7, in <module>pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16, revision="fp16")File "C:\conda\envs\ldm\lib\site-packages\diffusers\pipelines\pipeline_utils.py", line 1105, in from_pretrainedloaded_sub_model = load_sub_model(File "C:\conda\envs\ldm\lib\site-packages\diffusers\pipelines\pipeline_utils.py", line 472, in load_sub_modelloaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)File "C:\conda\envs\ldm\lib\site-packages\transformers\modeling_utils.py", line 3019, in from_pretrainedstate_dict = load_state_dict(resolved_archive_file)File "C:\conda\envs\ldm\lib\site-packages\transformers\modeling_utils.py", line 463, in load_state_dictwith safe_open(checkpoint_file, framework="pt") as f:
safetensors_rust.SafetensorError: Error while deserializing header: HeaderTooLargeProcess finished with exit code 1
出现原因:
HeaderTooLarge
说明头解析错误
- LFS在git下载中被略过,文件是空指针
- 文档损坏
- 不是safetensor格式的权重
- 权重读取时被修改
解决
- 从github或者hugging face原链接手动下载,注意标有LFS文件
- 使用flag,强制下载LFS
- 重新下载权重
- 不使用软连接
- 查看文件大小、检查md5 code