The model can be trained and inferenced successfully now: Set safe_serialization to False in model training file tutorial_train_plus.py:

accelerator.save_state(save_path, safe_serialization=False)

It will generate pytorch_model.bin instead of model.safetensors during training.

Once training is complete, modify the model conversion code as below based on the original instructions in readme:

ckpt = "pytorch_model.bin" # set correct path
sd = torch.load(ckpt)

Model file ip_adapter.bin will be generated for inference.