HCAttention: Extreme KV Cache Compression via Heterogeneous Attention Computing for LLMs
This addresses the problem of high memory requirements during inference for users of large language models, representing a strong specific gain rather than an incremental improvement.
The paper tackles the memory challenge of processing long-context inputs in large language models by proposing HCAttention, a heterogeneous attention framework that reduces the KV cache memory footprint to 25% of its original size while preserving accuracy, and extends the Llama-3-8B model to handle 4 million tokens on a single A100 GPU.
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.