---
title: 'Fast Transformer Decoding: One Write-Head is All You Need'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/1911-02150-multi-query-attention.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Noam Shazeer, Fast Transformer Decoding: One Write-Head is All You Need (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: efficiency'
tags: arxiv, efficiency, attention, inference
arxiv_id: 1911.02150
alias: 1911.02150-multi-query-attention
type: pointer
format: pointer
sources:
  - id: S1
    title: 'Fast Transformer Decoding: One Write-Head is All You Need'
    publisher: arXiv
    authors: 'Noam Shazeer'
    url: https://arxiv.org/abs/1911.02150
    version_read: 'v1 (6 Nov 2019)'
    accessed: 2026-11-12
    license: 'arXiv non-exclusive distribution license'
---
# Fast Transformer Decoding: One Write-Head is All You Need

## Summary
[Fast Transformer Decoding: One Write-Head is All You Need](https://arxiv.org/abs/1911.02150) (2019) introduced multi-query attention, in which many query heads share a single key and value head. The point is faster incremental decoding: inference is often limited by repeatedly loading a large key-value cache, and shrinking that cache reduces the memory-bandwidth cost.

What it introduced: the idea that you do not need a separate key and value projection for every attention head at decode time. One shared write-head for keys and values can preserve most of the quality of multi-head attention while making incremental inference much cheaper.

What it led to: grouped-query attention, now a default in many open language models, and the broader focus on KV-cache cost as a first-class inference constraint. Live note: [2305-13245-gqa](https://iinventedlight.com/library/notes/2305-13245-gqa/). This short paper is the ancestor of that design.

Use it when you want the original multi-query attention note. Where to read it: https://arxiv.org/abs/1911.02150 (v1 (6 Nov 2019)).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
