RestoreKV recovers KV cache performance via LoRA-adapted restoration
August 1, 2026
RestoreKV uses a small number of learned restore tokens and a LoRA-adapted pass to recover lost information in query-agnostic KV cache eviction. This allows for aggressive compression while maintaining the performance of a full-cache implementation.
HOW THIS AFFECTS YOU
●
builderYou can implement more aggressive KV cache compression to save memory without a major drop in performance.