ModuleNotFoundError: No module named 'nunchaku' in ComfyUI

For ModuleNotFoundError: No module named 'nunchaku' (and nunchaku.lora, nunchaku.models, or red NunchakuFluxLoraLoader / NunchakuFluxDiTLoader nodes): do not install the unrelated PyPI package named nunchaku — it is an unrelated scientific package. First confirm whether the ComfyUI-nunchaku plugin imported, then install the official Nunchaku wheel that matches your Python, PyTorch, CUDA, and GPU.

Exact error logs

ModuleNotFoundError: No module named 'nunchaku'
ModuleNotFoundError: No module named 'nunchaku.lora'
ModuleNotFoundError: No module named 'nunchaku.models'
ImportError: cannot import name 'SVDQW4A4Linear'
Node import failed: NunchakuFluxLoraLoader
Node import failed: NunchakuFluxLoraStack
Node import failed: NunchakuFluxDiTLoader

Red NunchakuFlux... node names are the visible workflow symptom. The root problem is usually that the ComfyUI-nunchaku plugin could not load its Nunchaku backend.

Nunchaku is a 4-bit quantized inference engine (SVDQuant method) used through the ComfyUI-nunchaku plugin to run quantized FLUX, Qwen-Image, SANA, and PixArt workflows on NVIDIA GPUs. ComfyUI runs fine without it — install it only when your workflow uses ComfyUI-nunchaku nodes or SVDQuant/Nunchaku quantized model files.

Triage

Check the environment before installing

Nunchaku wheels are tied to Python, PyTorch, CUDA, GPU architecture, and plugin version. Verify:

python --version
python -c "import torch; print(torch.__version__, torch.version.cuda)"

On the official Windows portable package, run from the portable root with .\python_embeded\python.exe -s .... Match the wheel to the Python tag (cp311, cp312), PyTorch version, CUDA version, Windows architecture, and GPU generation when the docs mention it. Nunchaku is primarily for NVIDIA GPUs; on AMD, Intel, Apple Silicon, or CPU-only hardware, use a different path such as GGUF quantization or a smaller model.

Installation methods

Method 1: ComfyUI-nunchaku installer workflow

If you have a recent ComfyUI-nunchaku plugin, use its install_wheel.json workflow (the official ComfyUI install path):

1. Load install_wheel.json in ComfyUI.
2. Run the installer node in update node mode to fetch available versions.
3. Select the wheel matching your Python, PyTorch, CUDA, and GPU.
4. Run the node in install mode.
5. Fully restart ComfyUI (a browser refresh is not enough — custom nodes register at backend startup).

Method 2: Manually install a matching wheel

Download a wheel from Nunchaku's official release sources (GitHub may redirect between org names):

• GitHub Releases: https://github.com/nunchaku-tech/nunchaku/releases
• Hugging Face: https://huggingface.co/nunchaku-tech
• ModelScope: https://modelscope.cn/organization/nunchaku-tech

Then install with the Python that starts ComfyUI:

python -m pip install <path-or-url-to-matching-nunchaku-wheel.whl>

For the official Windows portable package, use .\python_embeded\python.exe -s -m pip install <...>. Do not copy an example wheel URL blindly — a wheel for Python 3.11 / PyTorch 2.7 will not repair a Python 3.12 / different PyTorch environment.

Method 3: Build from source only if you need it

Building on Windows requires a compatible CUDA toolkit, Visual Studio/MSVC, and the correct dev environment. For most users, a prebuilt wheel or the installer workflow is the practical path.

Common installation failures

• No matching distribution found — your Python/PyTorch/CUDA combination has no matching wheel from that source. Re-check versions and pick a wheel built for that exact combination.
• Wrong package installed — if you installed the unrelated PyPI nunchaku, uninstall it (python -m pip uninstall nunchaku -y) before installing the correct wheel.
• Local directory conflict — a folder named nunchaku in the working directory can shadow the installed package. Remove any conflicting folder under the ComfyUI root or plugin folder.
• Node still red after the wheel imports — the backend may be installed while the custom node is stale or failed at startup. Check http://127.0.0.1:8188/v2/customnode/import_fail_info_bulk and look for ComfyUI-nunchaku. If an earlier dependency conflict stopped the plugin, repair that earliest failure first (see dependency conflicts).
• RTX 20 series — older Turing GPUs have different datatype/attention constraints than Ada/Blackwell. Follow the Nunchaku docs for node settings for your GPU generation.

Verify

python -c "import nunchaku; print(getattr(nunchaku, '__version__', 'installed'))"

Then fully restart ComfyUI and confirm the ComfyUI-nunchaku nodes register.

Alternatives

If your hardware does not meet Nunchaku's requirements: full-precision models (with enough VRAM), GGUF quantization (easier across setups), smaller models (SDXL/SD 1.5), or cloud GPU.

Source References

• ComfyUI-nunchaku GitHub repository
• ComfyUI-nunchaku installation docs
• Nunchaku installation docs
• Nunchaku FAQ: wrong PyPI package
• Nunchaku wheel repository on Hugging Face
• Nunchaku GitHub releases
• ComfyUI custom node troubleshooting guide