Nunchaku Installer

Here's the thing nobody tells you about Nunchaku: installing the ComfyUI nodes is only half the job. The nodes are just a front end. The actual 4-bit inference engine - the nunchaku Python wheel - is a separate compiled package, and until it's installed and matches your setup, every Nunchaku node either errors or refuses to appear. NunchakuWheelInstaller is the node the authors built to install that engine for you, so you don't have to hand-pick a wheel off a releases page.

If you searched this node's name, you were probably staring at a red error or a missing-node warning. This is the fix.

How it works

A Python wheel is a prebuilt binary, and Nunchaku's is compiled against a specific PyTorch version, Python version, and CUDA toolkit. Get any of those wrong and it won't import. The right wheel for a Torch 2.5 + Python 3.11 box is a different file from the one for Torch 2.6 + Python 3.12. Picking by hand means reading a compatibility table and not fat-fingering it.

This node reads your running environment, figures out which wheel you need, downloads it, and installs it into the same Python ComfyUI is using - which is the other classic trap, since pip install nunchaku in the wrong terminal lands the wheel in a venv ComfyUI never touches. Run this node from inside ComfyUI and it can't miss the environment. It was added back in v0.3.1 precisely because manual wheel installs were burning people.

The inputs and outputs that matter

• mode (install / uninstall / update node, default install) - the one you actually set. install puts the engine in; update node refreshes the ComfyUI plugin itself; uninstall removes the wheel. For a first run, leave it on install.
• version and dev_version - dropdowns for pinning a specific engine build. They read none until the node has populated the list of available wheels for your machine, at which point you pick a stable release (or a dev build if you're chasing a brand-new model). Most people take the newest stable and move on.

It outputs a status string - literally the install log/result - and it's an output node, so you queue the prompt and read what it says. "Success, restart ComfyUI" is what you want.

How to install it

Slight chicken-and-egg: this node ships with the ComfyUI-nunchaku pack, so you install the pack the normal way and this node comes along. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. Now drop a NunchakuWheelInstaller node on the canvas, set mode to install, queue it, read the status, and restart ComfyUI so the freshly installed engine gets picked up. After that the loader nodes will actually work.

Common issues & troubleshooting

You installed the pack and the loaders still error. That's the whole reason this node exists - the pack is nodes, not the engine. Run this, restart, done.

It says success but nothing changed. You didn't restart. The wheel loads at Python startup; ComfyUI has to come back up to see it.

Install fails outright. Usually a version mismatch it can't resolve, or no wheel exists yet for a bleeding-edge Torch/Python combo. Check the pin dropdowns for an available build, or fall back to the manual install from the Nunchaku docs - the README links the full compatibility guide.

On a hosted/managed ComfyUI where the Python environment is locked, an in-graph installer may not be able to write to it. There, the engine is (or has to be) baked into the image ahead of time rather than installed at runtime.