Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction
Fast KVzip uses lightweight sink-attention gates to evict up to 70% of KV caches with negligible overhead, maintaining near-lossless LLM performance across long-context, code, and math tasks via forward-only, task-agnostic training.
Published 2026Sydney Poster Session 3 · Wed, Dec 9, 10:00 AM–1:00 PM local time · Hall 1-4▲ 8 on Hugging FaceCode ★ 31arXiv ↗OpenReview ↗
Only vote on papers you've read. Sign in with GitHub to vote.
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.