Compactor: Calibrated Query-Agnostic KV Cache Compression with Approximate Leverage Scores

TL;DR

Compactor is a query-agnostic KV cache compression method for LLMs that uses approximate leverage scores to retain essential tokens, reducing memory usage by up to 68% while maintaining performance across diverse tasks.

Contribution

We introduce Compactor, a novel, training-free, query-agnostic compression strategy using leverage scores, with a context-calibrated procedure for optimal compression in LLMs.

Findings

Achieves 20% token retention with comparable performance to existing methods.

Reduces KV memory by 68% on Longbench with full performance.

Demonstrates effectiveness across 27 diverse tasks and models.

Abstract

Modern Large Language Models (LLMs) are increasingly trained to support very large context windows. We present Compactor, a training-free, query-agnostic KV compression strategy that uses approximate leverage scores to determine token importance. We show that Compactor can achieve the same performance as competing methods while retaining 20% fewer tokens in both synthetic and real-world context tasks, while being more task-robust. We further introduce a procedure for context-calibrated compression: inferring the maximum compression a given context supports before significant performance loss. Using context-calibrated compression, we show that Compactor achieves full KV performance on Longbench while reducing the KV memory burden by 68%, on average. To demonstrate the efficacy and generalizability of our approach, we apply Compactor to 27 synthetic and real-world tasks from RULER and Longbench, with models from both the Qwen 2.5 and Llama 3.1 families. Finally, we release compactor-vllm, an inference engine and suite of optimized Triton kernels designed to efficiently support the sparse, non-contiguous memory access patterns inherent to compressed KV caches. This work demonstrates that Compactor offers a practical, high-performance solution for alleviating the memory bottleneck in modern LLM deployment.