Math identity that applied at model-load time, makes MLA decode at 71,000 tok/s(efficientagent.substack.com)
1 pointby bnayak252 hours ago1 comment
  • 2 hours ago
    undefined