> But none of those, except for dex, has a solution for fusing kernels that rely on loops. The LV rewrite will. Some day, I'd like to have it target accelerato

> But none of those, except for dex, has a solution for fusing kernels that rely on loops. The LV rewrite will. Some day, I'd like to have it target accelerators, but unlike fusion, I've not actually put any research/engineering into it so can't make any promises. But my long term goal is that simple loops in -> optimized anything you want out. Enzyme also deserves a shout out for being able to generate reverse mode AD loops with mutation.