On the Efficacy of Eviction Policy for Key-Value Constrained Generative Language Model Inference

TL;DR

This paper evaluates eviction policies for key-value caches in large language models, identifies shortcomings in existing methods, and introduces RoCo, a new robust policy that improves efficiency in resource-limited environments.

Contribution

The paper introduces RoCo, a novel eviction policy based on temporal attention scores and robustness, addressing deficiencies in prior policies for key-value cache management.

Findings

RoCo outperforms existing eviction policies in experiments.

RoCo improves efficiency during model inference.

EasyKV software package facilitates key-value constrained generative inference.

Abstract

Despite the recent success associated with Large Language Models (LLMs), they are notably cost-prohibitive to deploy in resource-constrained environments due to their excessive memory and computational demands. In addition to model parameters, the key-value cache is also stored in GPU memory, growing linearly with batch size and sequence length. As a remedy, recent works have proposed various eviction policies for maintaining the overhead of key-value cache under a given budget. This paper embarks on the efficacy of existing eviction policies in terms of importance score calculation and eviction scope construction. We identify the deficiency of prior policies in these two aspects and introduce RoCo, a robust cache omission policy based on temporal attention scores and robustness measures. Extensive experimentation spanning prefilling and auto-regressive decoding stages validates the superiority of RoCo. Finally, we release EasyKV, a versatile software package dedicated to user-friendly key-value constrained generative inference. Code available at https://github.com/DRSY/EasyKV.