Fast transformer inference with Metal Performance Shaders

Seems like the Apple M1 Max is almost half as fast as a RTX 3090. Pretty cool!