It's also just inference, not training.