---
license: bsd-3-clause
tags:
- flash-attention
- windows
- pytorch
- cuda
- whl
language:
- en
---

# FlashAttention Windows Wheel

Unofficial Windows-compatible wheel of flash-attention for Windows<br>
**Python 3.13-3.14 versions only**.

## Overview

This repository provides **Windows-compatible wheels** for FlashAttention-2 that are not officially distributed.<br>
Pre-built version: **flash_attn 2.9.2.post2** with **Python 3.13-3.14** support.

# !!Important!!

If you intend to use this alongside xformers, please note that versions v2.9.0 and later—which I developed myself—are not compatible. If you are using xformers, please use the latest official build, v2.8.4. I'll upload them.

## Key Features

- ✅ Native Windows support (Python 3.12-3.14)
- ⚡ FlashAttention-2

## Changelog

- 15.11.2025 Uploaded v2.8.3 based on PyTorch 2.9.1+cu130
- 12.02.2026 Uploaded v2.8.3 based on PyTorch 2.10.0+cu130
- 29.03.2026 Uploaded v2.8.3 based on PyTorch 2.11.0+cu130
- 14.05.2026 Uploaded v2.8.4 based on PyTorch 2.11.0+cu130
- 13.05.2026 Uploaded **v2.9.0** based on PyTorch 2.11.0+cu130 — **unofficial fork-only build** (not the official FlashAttention release). Includes FA2 A-1/A-2 optimizations.
- 15.05.2026 Uploaded **v2.9.0** based on PyTorch 2.12.0+cu132
- 23.05.2026 Uploaded **v2.9.1** based on PyTorch 2.12.0+cu132 — **unofficial fork-only build** (not the official FlashAttention release). Includes FA2 A-1/A-2 optimizations.
- 03.07.2026 Uploaded v2.8.4 based on Pytorch 2.12.1+cu132 **(Built on official-code for xformers-users)**
- 11.07.2026 Uploaded **v2.9.1** based on PyTorch 2.13.0+cu132 — **unofficial fork-only build** (not the official FlashAttention release). Includes FA2 A-1/A-2 optimizations.
- 12.07.2026 Uploaded **v2.9.2** based on PyTorch 2.13.0+cu132 — **unofficial fork-only build** (not the official FlashAttention release.Blackwell=SM100,120,121.Legacy=SM80,89,90). Includes FA2 A-1/A-2 optimizations.
- 18.07.2026 Uploaded v2.8.4 based on Pytorch 2.13.0+cu132 **(Built on official-code for xformers-users)**
- 12.08.2026 Uploaded **v2.9.2.post1** based on PyTorch 2.13.0+cu132 — **unofficial fork-only build** (not the official FlashAttention release.Blackwell=SM100,120,121.Legacy=SM80,89,90). Includes FA2 A-1/A-2 optimizations.
- 06.09.2026 Uploaded **v2.9.2.post2** based on PyTorch 2.14.0+cu132 — **unofficial fork-only build** (not the official FlashAttention release.Blackwell=SM100,120,121.Legacy=SM80,89,90). Includes FA2 A-1/A-2 optimizations.
- 06.09.2026 Uploaded v2.8.4 based on Pytorch 2.14.0+cu132 **(Built on official-code for xformers-users.Blackwell=SM100,120,121.Legacy=SM80,89,90)**

## About v2.9.x

**v2.9.x is not an official FlashAttention release.**<br>
It is an independent fork build that continues FA2 kernel development while upstream focuses on FA3/F4.

- **Optimization plan (GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/AI/FA2_BACKPORT_FROM_FA4_PLAN.md
- **Kernel change notes (GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/md/FA2_CHANGES_v1.2.md
- **About v2.9.1 A-1 / A-2 Refinements (GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/AI/A1_A2_REFINEMENTS_PLAN.md 
- **About v2.9.2 added 24 logics based on official repository (GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/md/split_align_kernels_explanation.md 
- **About v2.9.2.post1 A-1 / A-2 Refinements (GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/AI/A1_A2_REFINEMENTS_PLAN.md
- **About v2.9.2.post2 (GitHub):** https://github.com/ussoewwin/flash-attention/releases/tag/v1.8.1

## Disclaimer

- **No performance benchmarks** have been run on this build.
- **No multi-environment testing** has been performed.
  (However, It's Working on my Stable Diffusuion A1111(※Customed), Forge Nunchaku(※Customed) and ComfyUI)
- **Fork `flash_attn` 2.9.2.post2 — test of whether features work and can be used normally (log, GitHub):** https://github.com/ussoewwin/flash-attention/blob/main/md/2.9.2.post2_COMPLETE_TEST_AND_VALIDATION_GUIDE.md
- This is an **unofficial fork build**. Use at your own risk.

※Unofficial built version!! It works correctly in my environment, but I am not sure that will work in yours.