Flash Attention 3 - Precompiled Wheel

A precompiled wheel for Flash Attention 3 (v3.0.0). You can install it directly via pip

Compatibility

To use this wheel, your environment must match the following exact specifications:

• GPU: NVIDIA Hopper Architecture ONLY (H100, H200, GH200) sm_90a
• OS: Linux (x86_64)
• Python: 3.9, 3.10, 3.11, 3.12, or newer (abi3 compatible)
• PyTorch: 2.9.x
• CUDA: 12.8

How to Install

Download the .whl file from the assets below, or install it directly via URL:

# Example pip install
pip install https://github.com/aw920h/flash-attn-3-wheels/releases/download/v3.0.0-torch2.9.1-cu128/flash_attn_3-3.0.0-cp39-abi3-linux_x86_64.whl

Build Environment Details

• Base Repo: Official Flash Attention hopper branch
• Torch Version: 2.9.1+cu128
• NVCC Version: Built using internal PTXAS 12.8.93