KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing

TL;DR

KVSharer is a novel method that reduces memory and computation in large language model inference by sharing dissimilar key-value caches across layers, achieving significant efficiency gains without major performance loss.

Contribution

It introduces a layer-wise KV cache sharing technique that surprisingly benefits from dissimilar cache sharing, and demonstrates compatibility with existing intra-layer compression methods.

Findings

Reduces KV cache memory by 30% during inference.

Achieves at least 1.3x faster generation.

Compatible with existing intra-layer compression methods.

Abstract

The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \textit{KVSharer} can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.