LoRA output: tensor([[0.6639, 0.4487]], grad_fn=<AddmmBackward0>)