The checkpoints uploaded on the Hub use torch_dtype = 'float16', which will be
used by the AutoModel API to cast the checkpoints from torch.float32 to torch.float16.