Beyond Homogeneous Attention: Memory-Efficient LLMs via Fourier-Approximated KV Cache

TL;DR

FourierAttention introduces a memory-efficient, training-free method for large language models that uses Fourier basis projections to better handle long-context dependencies, improving accuracy and deployment efficiency.

Contribution

The paper presents FourierAttention, a novel approach exploiting heterogeneous transformer head roles with Fourier basis projections, enhancing long-context modeling without training or accuracy loss.

Findings

Achieves state-of-the-art long-context accuracy on benchmarks.

Enables efficient deployment with a custom Triton kernel.

Maintains performance while reducing memory usage.

Abstract

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.