Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction

TL;DR

Fast KVzip introduces a gating-based KV cache eviction method for LLMs that achieves high compression with minimal performance loss, enabling efficient inference across diverse tasks.

Contribution

It presents a novel, lightweight gating mechanism for KV cache eviction in frozen LLMs that maintains performance while significantly reducing cache size.

Findings

Up to 70% KV cache eviction with negligible performance loss.

Effective across multiple LLMs and tasks including reasoning and code comprehension.

Seamless integration into existing LLM inference pipelines.

Abstract

Efficient key-value (KV) cache management is crucial for the practical deployment of large language models (LLMs), yet existing compression techniques often incur a trade-off between performance degradation and computational overhead. We propose a novel gating-based KV cache eviction method for frozen-weight LLMs that achieves high compression ratios with negligible computational cost. Our approach introduces lightweight sink-attention gating modules to identify and retain critical KV pairs, and integrates seamlessly into both the prefill and decoding stages. The proposed gate training algorithm relies on forward passes of an LLM, avoiding expensive backpropagation, while achieving strong task generalization through a task-agnostic reconstruction objective. Extensive experiments across the Qwen2.5-1M, Qwen3, and Gemma3 families show that our method maintains near-lossless performance while evicting up to 70% of the KV cache. The results are consistent across a wide range of tasks, including long-context understanding, code comprehension, and mathematical reasoning, demonstrating the generality of our approach.