3 ms·

speed up transformer inference by 60%