Issues / #3149

#3149 support for q5 hadamard rotation for K or V or both

open · @Schumi08 · 0 comments · View on GitHub

Description

support for q5 hadamard rotation for K or V or both, will make them nearly lossless even at 256 ctx... with a small cost of compute, it will allow massive increase in ctx reducing KV size

Related on this site

Quick links to related pages.