Math identity that applied at model-load time, makes MLA decode at 71,000 tok/s