KVzip: Query-Agnostic KV Cache Compression with Context Reconstruction
This addresses memory and latency issues for LLM inference, particularly with long contexts, offering a practical improvement over existing methods.
The paper tackles the problem of memory overhead and latency from large KV caches in LLMs by introducing KVzip, a query-agnostic compression method that reduces KV cache size by 3-4x and decoding latency by about 2x with minimal performance loss across various tasks.
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.