0

Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning

HeadKV and HeadKV-R2 improve efficiency in Large Language Models by compressing the KV cache at the head level, significantly outperforming baselines on contextual QA tasks with limited resources.

Year
2024
Venue
arXiv 2024
Authors
6
Hosting
Abstract onlyARXIV-DEFAULT

Cite

Notes

Only stored in your browser.

Attribution

Abstract & full text
arxiv.org/abs/2410.19258v3ARXIV-DEFAULT
TL;DR
Semantic Scholar
Attribution policy →

Abstract

Key-Value (KV) caching is a common technique to enhance the computational efficiency of Large Language Models (LLMs), but its memory overhead grows rapidly with input length. Prior work has shown that not all tokens are equally important for text generation, proposing layer-level KV cache compression to selectively retain key information. Recognizing the distinct roles of attention heads in generation, we propose HeadKV, a head-level KV cache compression method, and HeadKV-R2, which leverages a novel contextual reasoning ability estimation for compression. Our approach operates at the level of individual heads, estimating their importance for contextual QA tasks that require both retrieval and reasoning capabilities. Extensive experiments across diverse benchmarks (LongBench, LooGLE), model architectures (e.g., Llama-3-8B-Instruct, Mistral-7B-Instruct), and long-context abilities tests demonstrate that our head-level KV cache compression significantly outperforms strong baselines, particularly in low-resource settings (KV size = 64 & 128). Notably, our method retains just 1.5% of the KV cache while achieving 97% of the performance of the full KV cache on the contextual question answering benchmark.Codes are available at https://github.com/FYYFU/HeadKV

Authors

6