HCAttention: Extreme KV Cache Compression via Heterogeneous Attention Computing for LLMs

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Yang, Dongquan, Yang, Yifan, Yu, Xiaotian, Qi, Xianbiao, Xiao, Rong
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866911078393315328
author Yang, Dongquan
Yang, Yifan
Yu, Xiaotian
Qi, Xianbiao
Xiao, Rong
author_facet Yang, Dongquan
Yang, Yifan
Yu, Xiaotian
Qi, Xianbiao
Xiao, Rong
contents Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.
format Preprint
id arxiv_https___arxiv_org_abs_2507_19823
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle HCAttention: Extreme KV Cache Compression via Heterogeneous Attention Computing for LLMs
Yang, Dongquan
Yang, Yifan
Yu, Xiaotian
Qi, Xianbiao
Xiao, Rong
Computation and Language
Artificial Intelligence
Machine Learning
Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.
title HCAttention: Extreme KV Cache Compression via Heterogeneous Attention Computing for LLMs
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2507.19823