DuoAttention-Slashes memory and latency for LLMs without sacrificing performance | Hacker News Reader