Multi-Head Low-Rank Attention

TL;DR

This paper introduces Multi-Head Low-Rank Attention (MLRA), a novel method that improves large language model decoding efficiency by enabling partitionable latent states, reducing memory traffic, and achieving faster inference without sacrificing performance.

Contribution

MLRA provides a partitionable latent state mechanism for efficient distributed decoding, overcoming sharding bottlenecks of prior methods like MLA.

Findings

Achieves state-of-the-art perplexity and downstream task performance.

Delivers a 2.8× decoding speedup over MLA.

Reduces memory traffic during distributed decoding.

Abstract

Long-context inference in large language models is bottlenecked by Key--Value (KV) cache loading during the decoding stage, where the sequential nature of generation requires repeatedly transferring the KV cache from off-chip High-Bandwidth Memory (HBM) to on-chip Static Random-Access Memory (SRAM) at each step. While Multi-Head Latent Attention (MLA) significantly reduces the total KV cache size, it suffers from a sharding bottleneck during distributed decoding via Tensor Parallelism (TP). Since its single latent head cannot be partitioned, each device is forced to redundantly load the complete KV cache for every token, consuming excessive memory traffic and diminishing TP benefits like weight sharding. In this work, we propose Multi-Head Low-Rank Attention (MLRA), which enables partitionable latent states for efficient 4-way TP decoding. Extensive experiments show that MLRA achieves state-of-the-art perplexity and downstream task performance, while also delivering a 2.8$\times$ decoding speedup over MLA. Code is available at https://github.com/SongtaoLiu0823/MLRA. Pretrained weights, along with the training and evaluation data, are available at https://huggingface.co/Soughing/MLRA.