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

 (efficientagent.substack.com)

2 points | by bnayak25 4 hours ago ago

1 comments