Y Hacker News new | ask | show | jobs

user: thw20 created: 2024-04-24 karma: 1 submissions: Simple, zero overhead way to compress model, KV cache via Low-Rank Decomposition 1 points | 0 comments 0 points | 0 comments 0 points | 0 comments 0 points | 0 comments Towards understanding multiple attention sinks in LLMs 1 points | 2 comments 0 points | 0 comments The Existence and Behavior of Secondary Attention Sinks 1 points | 0 comments