KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Kim, Jang-Hyun, Kim, Jinuk, Kwon, Sangwoo, Lee, Jae W., Yun, Sangdoo, Song, Hyun Oh
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909816118575104
author Kim, Jang-Hyun
Kim, Jinuk
Kwon, Sangwoo
Lee, Jae W.
Yun, Sangdoo
Song, Hyun Oh
author_facet Kim, Jang-Hyun
Kim, Jinuk
Kwon, Sangwoo
Lee, Jae W.
Yun, Sangdoo
Song, Hyun Oh
contents Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by $3$-$4\times$ and FlashAttention decoding latency by approximately $2\times$, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1, Qwen2.5, and Gemma3, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90% cache budget ratio under multi-query scenarios.
format Preprint
id arxiv_https___arxiv_org_abs_2505_23416
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
Kim, Jang-Hyun
Kim, Jinuk
Kwon, Sangwoo
Lee, Jae W.
Yun, Sangdoo
Song, Hyun Oh
Databases
Machine Learning
Transformer-based large language models (LLMs) cache context as key-value (KV) pairs during inference. As context length grows, KV cache sizes expand, leading to substantial memory overhead and increased attention latency. This paper introduces KVzip, a query-agnostic KV cache eviction method enabling effective reuse of compressed KV caches across diverse queries. KVzip quantifies the importance of a KV pair using the underlying LLM to reconstruct original contexts from cached KV pairs, subsequently evicting pairs with lower importance. Extensive empirical evaluations demonstrate that KVzip reduces KV cache size by $3$-$4\times$ and FlashAttention decoding latency by approximately $2\times$, with negligible performance loss in question-answering, retrieval, reasoning, and code comprehension tasks. Evaluations include various models such as LLaMA3.1, Qwen2.5, and Gemma3, with context lengths reaching up to 170K tokens. KVzip significantly outperforms existing query-aware KV eviction methods, which suffer from performance degradation even at a 90% cache budget ratio under multi-query scenarios.
title KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
topic Databases
Machine Learning
url https://arxiv.org/abs/2505.23416