# SageAttention Wheels (CUDA 13.x) 🚀

Prebuilt SageAttention 2.2.0 wheels compiled for Linux x86_64 with CUDA 13.x support.

This repository provides ready-to-use binary wheels for different Python and PyTorch versions, optimized for modern NVIDIA GPUs (Ada / Hopper / Ampere).

---

# 📦 Available Wheels

| Python Version | PyTorch Version | CUDA | File |
|----------------|----------------|------|------|
| 3.11 | 2.10 | cu13 | sageattention-2.2.0-python3.11-pytorch2.10-cu13-linux_x86_64.whl |
| 3.12 | 2.10 | cu13 | sageattention-2.2.0-python3.12-pytorch2.10-cu13-linux_x86_64.whl |
| 3.12 | 2.11 | cu13 | sageattention-2.2.0-python3.12-pytorch2.11-cu13-linux_x86_64.whl |
| 3.13 | 2.11 | cu13 | sageattention-2.2.0-python3.13-pytorch2.11-cu13-linux_x86_64.whl |

---

# ⚡ Requirements

- Linux x86_64
- NVIDIA GPU (Ada / Ampere / Hopper tested)
- CUDA 13.x runtime / toolkit
- PyTorch matching wheel version
- Python version matching wheel

---

# 🧠 Installation

## 1. Create virtual environment

python3.12 -m venv venv
source venv/bin/activate

---

## 2. Install PyTorch (CUDA 13)

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

---

## 3. Install SageAttention wheel

pip install sageattention-2.2.0-python3.12-pytorch2.11-cu13-linux_x86_64.whl

---

# 🧪 Quick Test

python -c "import torch; print(torch.cuda.is_available())"
python -c "import sageattention; print('SageAttention loaded successfully')"

---

# 🚀 Notes

- Wheels are precompiled for performance
- Must match Python + PyTorch versions exactly
- CUDA 13.x required
- Optimized for sm_80+ GPUs

---

# ⚠️ Troubleshooting

CUDA not found:
export CUDA_HOME=/opt/cuda
export PATH=$CUDA_HOME/bin:$PATH
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

---

# 💬 Support
Matrix Network 

@aimiko:mochiart.moe

---

# 📜 License

Refer to upstream SageAttention repository.
This repo contains only prebuilt binaries.