I used the updated last step to test the fine-tuned model. It answers correctly but it doesn't seem to use the GPU.