Fast Transformer Decoding: One Write-Head is All You Need

TL;DR

This paper introduces multi-query attention, a variant of multi-head attention in Transformers, which shares keys and values across heads to significantly speed up incremental decoding with minimal quality loss.

Contribution

The paper proposes multi-query attention, reducing memory bandwidth during decoding and enabling faster inference without substantial accuracy compromise.

Findings

Decoding speed is significantly improved.

Minor quality degradation observed.

Memory bandwidth requirements are greatly reduced.

Abstract

Multi-head attention layers, as used in the Transformer neural sequence model, are a powerful alternative to RNNs for moving information across and between sequences. While training these layers is generally fast and simple, due to parallelizability across the length of the sequence, incremental inference (where such paralleization is impossible) is often slow, due to the memory-bandwidth cost of repeatedly loading the large "keys" and "values" tensors. We propose a variant called multi-query attention, where the keys and values are shared across all of the different attention "heads", greatly reducing the size of these tensors and hence the memory bandwidth requirements of incremental decoding. We verify experimentally that the resulting models can indeed be much faster to decode, and incur only minor quality degradation from the baseline.