well, it's compute optimal for some inference setups, I think that's the idea; it's not optimal for training though

→ View original post on X — @jxmnop