Beyond Homogeneous Attention: Memory-Efficient LLMs via Fourier-Approximated KV Cache
This addresses memory efficiency for LLM deployment, particularly for long-context applications, though it is an incremental improvement over existing compression methods.
The paper tackles the memory demands of Key-Value (KV) caches in Large Language Models by proposing FourierAttention, a training-free framework that approximates long-context-insensitive dimensions using Fourier bases, achieving the best long-context accuracy on LongBench and Needle-In-A-Haystack benchmarks.
Large Language Models struggle with memory demands from the growing Key-Value (KV) cache as context lengths increase. Existing compression methods homogenize head dimensions or rely on attention-guided token pruning, often sacrificing accuracy or introducing computational overhead. We propose FourierAttention, a training-free framework that exploits the heterogeneous roles of transformer head dimensions: lower dimensions prioritize local context, while upper ones capture long-range dependencies. By projecting the long-context-insensitive dimensions onto orthogonal Fourier bases, FourierAttention approximates their temporal evolution with fixed-length spectral coefficients. Evaluations on LLaMA models show that FourierAttention achieves the best long-context accuracy on LongBench and Needle-In-A-Haystack (NIAH). Besides, a custom Triton kernel, FlashFourierAttention, is designed to optimize memory via streamlined read-write operations, enabling efficient deployment without performance compromise.