Flash-Attention 3 Wheels

Pre-built wheels that erase Flash Attention 3 installation headaches — now with Windows & Arm support! 🎉

**🚀 Update: We've successfully built Flash Attention 3 wheels for Windows and Arm CUDA SBSA platforms like GH200!

Upstream PR: Windows compatibility fixes submitted to Dao-AILab/flash-attention#2047

Recent Updates

• 2026-03-19: 🆕 CUDA 13.0 support added for Windows and Linux
• 2026-03-16: ⚙️ GitHub Actions updated to Node.js 24 compatible versions
• 2026-03-04: 🐛 Fixed corrupt patch error in Windows builds
• 2026-03-04: 🔧 Build scripts now clone directly from upstream flash-attention
• Earlier: 🪟 Windows and Arm64 (GH200) wheel support added; upstream PR #2047 submitted

Quick start

Pick the line that matches your setup (change cu128 / torch280 if needed):

# CUDA 12.8 + PyTorch 2.8.0
pip install flash_attn_3 \
  --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch280

How to pick the right index

Visit the GitHub Pages site and choose the link that matches:

• CUDA 13.0 → cu130_torch...
• CUDA 12.9 → cu129_torch...
• CUDA 12.8 → cu128_torch...
• CUDA 12.6 → cu126_torch...

Each page shows the one-liner you need.

When are wheels updated?

• Biweekly, every 2nd and 4th Sunday at 22:00 UTC
• On demand, by triggering the workflow manually if you need a fresher build

Releases are tagged with the build date (2025.10.15) so you always know how fresh your wheel is.

License

The build scripts and index generator are Apache-2.0.