Skipping 90% of KV dequant work speeds up LLM decode by 22%(github.com)1 points·by pidtom·4 maanden geleden·0 commentsgithub.comSkipping 90% of KV dequant work speeds up LLM decode by 22%https://github.com/TheTom/turboquant_plus1 commentsPost comment[+]pidtom·4 maanden geleden