想要从timm加载本地预训练模型，首先是参考timm.create_model()从本地加载pretrained模型
将

model = timm.create_model('modelxxx', pretrained=True,  xxx)

改为

pretrained_cfg = timm.models.create_model("modelxxx").default_cfg
pretrained_cfg['file'] = 'path/to/checkpoint'
model = timm.models.create_model("modelxxx", pretrained=True, xxx, pretrained_cfg=pretrained_cfg))  

但是遇到错误

AssertionError: pretrained_cfg should not be set when sourcing model from Hugging Face Hub.

后来参考LocalEntryNotFoundError when loading downloaded pretrained model using timm.create_model (side load offline weights, e.g. on Kaggle) #1826
成功将从本地加载预训练模型

timm.create_model(
  'modelxxx',
  pretrained=True,
  pretrained_cfg_overlay=dict(file='path/to/checkpoint'),
)

---------------------xxx-----------------------
从本地加载

model = timm.create_model('modelxxx', pretrained=False,  xxx)    # pretrained=True —> False
model.load_state_dict(torch.load(pth_local_path), strict=True)