gdn

LinearKV is a training-free framework that enables position-independent caching for hybrid LLMs by using a single cached state to initialize linear layers, outperforming exact prefix-state composition and remaining compatible with existing PIC methods.