4 ms·

This is incredibly useful. Thanks for making the kernels public. I'm curious if anyone has tried generalizing this to batched matmuls or to sparse inputs on Ad