JusteLeo/Nunchaku-Zimage-Win-Wheels

Nunchaku Prebuilt Wheels for Z-image (Windows)

This repository provides the missing pre-compiled binaries (.whl) for the Nunchaku library, specifically optimized for Windows x64 environments where official wheels are unavailable.

🚀 What's New in v1.2.0

• Cache-DiT: Support for the new inference optimization.
• Z-Image Turbo Optimizations: Performance gains via layer fusion (QKV/Norm/Rotary).
• RTX 2000 Series Support: Native compatibility for Turing GPUs.
• Improved Compatibility: Updated for the latest diffusers versions.

---

🛠 GPU & CUDA Support

⚠️ Hardware Compatibility Note

Version	Supported GPU Architectures
v1.2.0 (Latest)	RTX 2000, 3000, 4000, 5000 (Turing to Blackwell)
v1.1.0 (Legacy)	RTX 3000, 4000 (Ampere & Ada Lovelace only)

⚙️ Compilation Details

To ensure maximum stability and performance, the v1.2.0 wheels were compiled using CUDA Toolkit 12.9 for the following environments:

• PyTorch 2.7.0 +cu128
• PyTorch 2.8.0 +cu128

---

Compatibility Matrix (Windows Only)

Python Version	Torch 2.7.0 (cu128)	Torch 2.8.0 (cu128)
Python 3.11	✅ Available	✅ Available
Python 3.12	✅ Available	✅ Available
Python 3.13	✅ Available	✅ Available

---

Build Methodology

Built in 2025 with:

• OS: Windows 11 x64
• Compiler: MSVC (Visual Studio 2022)
• CUDA Toolkit: v12.9

Local Build Command (Reference)

set CUDA_HOME=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9
set CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.9
set DISTUTILS_USE_SDK=1

uv pip install torch==2.8.0 torchvision --index-url https://download.pytorch.org/whl/cu128
uv pip install numpy ninja setuptools packaging wheel
uv build --wheel --no-build-isolation

---

Installation

1. Download the .whl file matching your environment from the Files tab.
2. Install via pip:

# Example for Python 3.12 and v1.2.0
pip install nunchaku-1.2.0+torch2.8-cp312-cp312-win_amd64.whl

3. Verify:

import nunchaku
print("Nunchaku v1.2.0 (Z-image) successfully loaded.")

---

Disclaimer

This is an unofficial community repository providing missing Windows wheels. For original source code and official updates, please visit the Nunchaku-tech GitHub.