Essential Wheels (Windows)

Precompiled Windows wheels for rare Python / PyTorch / CUDA combinations that are no longer commonly supported by upstream projects.

This repository focuses on Python 3.10 (cp310), PyTorch 2.9.1, and CUDA 13.0 (cu130) — combinations that are increasingly hard to find prebuilt binaries for.

Important, you'll need PyTorch 2.9.1 CUDA 13.0, install command :

pip install torch==2.9.1 torchvision==0.24.1 torchaudio==2.9.1 --index-url https://download.pytorch.org/whl/cu130

Useful tips (optional, if needed) : You can install compatible triton-windows using this command :

pip install triton-windows==3.5.1.post24

You can install compatible xformers using this command :

pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu130

---

🔧 What’s included

This repository currently provides precompiled Windows wheels for:

• flash-attn 2.8.3
• sageattention 2.2.0
• llama-cpp-python 0.3.16

All wheels:

• Target Windows x86_64
• Are built for Python 3.10 (cp310)
• Expect PyTorch 2.9.1
• Target CUDA 13.0 (cu130)
• Built and validated on RTX 5090 / RTX 5080
• Expected to be backwards compatible with earlier NVIDIA GPUs (driver/runtime permitting)

---

📦 Installation

Option A — Install from a downloaded wheel (recommended)

1. Go to the Releases page of this repository
2. Download the .whl file you need
3. Install it locally:

pip install flash_attn-2.8.3-cp310-cp310-win_amd64.whl
pip install sageattention-2.2.0-cp310-cp310-win_amd64.whl
pip install llama_cpp_python-0.3.16-cp310-cp310-win_amd64.whl

Option B — Install directly from GitHub (advanced) : Flash attention :

pip install "https://github.com/MarkOrez/essential-wheels/releases/download/flash_attn-2.8.3-cp310-cp310-win_amd64.whl"

Sage attention :

pip install "https://github.com/MarkOrez/essential-wheels/releases/download/sageattention-2.2.0-cp310-cp310-win_amd64.whl"

Llama cpp python :

pip install "https://github.com/MarkOrez/essential-wheels/releases/download/llama_cpp_python-0.3.16-cp310-cp310-win_amd64.whl"

⚠️ Compatibility notes

These wheels are not official upstream builds

They are provided for convenience for users stuck on specific environments

CUDA / driver compatibility depends on your local system

📜 Licensing & attribution

Each wheel is built from its respective upstream project. Please refer to the original repositories for their licenses and source code:

FlashAttention

SageAttention

llama-cpp-python

This repository redistributes binaries for convenience only.