CORM: Cache Optimization with Recent Message for Large Language Model Inference

TL;DR

This paper introduces CORM, a cache optimization method for large language model inference that significantly reduces KV cache memory usage by up to 70% without notable performance loss, enabling more efficient deployment.

Contribution

CORM is a novel cache eviction policy leveraging token similarity and attention dependencies, reducing memory footprint without requiring model fine-tuning.

Findings

KV cache memory reduced by up to 70%

Negligible performance degradation across six tasks

Compatible with GQA for additional compression

Abstract

Large Language Models (LLMs), despite their remarkable performance across a wide range of tasks, necessitate substantial GPU memory and consume significant computational resources. Beyond the memory taken up by model weights, the memory used by the KV cache rises linearly with sequence length, becoming a primary bottleneck for inference. In this paper, we introduce an innovative method for optimizing the KV cache, which considerably minimizes its memory footprint. Upon thorough investigation, we discover that in most Transformer models, (i) there is a striking similarity between adjacent tokens' query vectors, and (ii) the attention calculation of the current query can rely exclusively on the attention information of a small fraction of preceding queries. Based on these observations, we present CORM, a KV cache eviction policy that dynamically retains essential key-value pairs for inference without the need for model fine-tuning. Our validation shows that CORM reduces the inference memory usage of KV cache by up to 70\% with negligible performance degradation across six tasks in LongBench. Furthermore, we demonstrate that CORM is compatible with GQA for further compression rate.