MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference

TL;DR

MadaKV introduces a dynamic, modality-aware cache eviction strategy that significantly reduces memory and latency in multimodal large language models during long-context inference, while maintaining high accuracy.

Contribution

It proposes a novel, adaptive KV cache eviction method tailored for multimodal models, addressing modality importance disparities across attention heads.

Findings

Reduces KV cache memory footprint substantially.

Improves inference decoding latency by 1.3 to 1.5 times.

Maintains high accuracy across various multimodal tasks.

Abstract

This paper introduces MadaKV, a modality-adaptive key-value (KV) cache eviction strategy designed to enhance the efficiency of multimodal large language models (MLLMs) in long-context inference. In multimodal scenarios, attention heads exhibit varying preferences for different modalities, resulting in significant disparities in modality importance across attention heads. Traditional KV cache eviction methods, which are tailored for unimodal settings, fail to capture modality-specific information, thereby yielding suboptimal performance. MadaKV addresses these challenges through two key components: modality preference adaptation and hierarchical compression compensation. By dynamically sensing modality information within attention heads and adaptively retaining critical tokens, MadaKV achieves substantial reductions in KV cache memory footprint and model inference decoding latency (1.3 to 1.5 times improvement) while maintaining high accuracy across various multimodal long-context tasks. Extensive experiments on representative MLLMs and the MileBench benchmark demonstrate the effectiveness of MadaKV compared to existing KV cache eviction methods.