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

TL;DR

HCAttention introduces a heterogeneous attention framework that significantly compresses KV cache memory in large language models, enabling long-context processing with minimal accuracy loss and without model fine-tuning.

Contribution

It is the first to combine key quantization, value offloading, and dynamic eviction for extreme KV cache compression compatible with existing transformers.

Findings

Preserves full-attention accuracy with only 25% KV cache memory.

Achieves state-of-the-art compression, using only 12.5% of cache.

Enables processing 4 million tokens on a single GPU with 80GB memory.

Abstract

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.