speed up transformer inference by 60%