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

(efficientagent.substack.com)

2 points | by bnayak25 5 hours ago

1 comments