SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference

TL;DR

SmallKV introduces a novel approach using a small model to compensate for KV cache compression in LLMs, addressing eviction issues and improving inference efficiency in resource-limited settings.

Contribution

The paper proposes SmallKV, a new method leveraging a small model to maintain attention alignment and improve KV cache management during LLM inference.

Findings

Achieves 1.75-2.56x higher throughput than baseline methods.

Effectively maintains attention matching across different-scale LLMs.

Demonstrates improved performance on multiple benchmark datasets.

Abstract

KV cache eviction has emerged as an effective solution to alleviate resource constraints faced by LLMs in long-context scenarios. However, existing token-level eviction methods often overlook two critical aspects: (1) their irreversible eviction strategy fails to adapt to dynamic attention patterns during decoding (the saliency shift problem), and (2) they treat both marginally important tokens and truly unimportant tokens equally, despite the collective significance of marginal tokens to model performance (the marginal information over-compression problem). To address these issues, we design two compensation mechanisms based on the high similarity of attention matrices between LLMs of different scales. We propose SmallKV, a small model assisted compensation method for KV cache compression. SmallKV can maintain attention matching between different-scale LLMs to: 1) assist the larger model in perceiving globally important information of attention; and 2) use the smaller model's attention scores to approximate those of marginal tokens in the larger model. Extensive experiments on benchmarks including GSM8K, BBH, MT-Bench, and LongBench demonstrate the effectiveness of SmallKV. Moreover, efficiency evaluations show that SmallKV achieves 1.75 - 2.56 times higher throughput than baseline methods, highlighting its potential for efficient and performant LLM inference in resource constrained environments.