@wyli Do you think you can help me with this error in the github action? I can't reproduce it locally. I think it's complaining that img_.device doesn't work when img_ is a np.ndarray. But the code before looks like this:

if get_track_meta():
    img_ = img if isinstance(img, MetaTensor) else MetaTensor(img)
    img_, *_ = convert_data_type(img, dtype=_dtype)
else:
    img_, *_ = convert_data_type(img, torch.Tensor, dtype=_dtype)

It seems to me that in all cases, img_ should be of type torch.Tensor and so `device`` should be available.

img_ should be of type torch.Tensor and so `device`` should be available.

looks like you addressed the issue... I'll merge this PR and #4308 and have the full integration tests soon.