🔧 Nunchaku Installer

Nunchaku is the MIT research project that runs FLUX at INT4 precision - the SVDQuant work that lets a quantized FLUX dev model sit in roughly 6.6GB instead of 23GB, and a big deal on the RTX 50-series where the older kernels and PyTorch builds don't play nicely. Installing it is normally a chore of matching CUDA and PyTorch and Python versions by hand. ArchAi3D Nunchaku Installer automates the matching: you pick an action, it detects your GPU, CUDA, PyTorch, and Python, and installs the right wheel from the project's releases.

The whole node is one dropdown, action:

• Check System - reports what it detected (GPU architecture, CUDA, PyTorch, Python) without installing anything. This is the button to hit first; it's free and tells you whether the install will even succeed.
• Install Nunchaku - downloads and installs the matching wheel, then runs the project's config scripts.
• Update Versions File - refreshes the list of available Nunchaku releases the node checks against.
• Upgrade PyTorch (Blackwell) - installs the cu128 PyTorch build for RTX 5090/5080-class cards, which need a newer PyTorch than most ComfyUI installs ship with.
• Build xformers (Blackwell) - compiles xformers for the same Blackwell target.

force_pytorch_version and force_python_version override auto-detection if the node guesses wrong. Output is a single status string - read it, because this node does real things to your environment.

The honest warning

This is an installer node: it runs pip installs (and on the "Build" action, a compile) into whatever Python environment ComfyUI is using. That's exactly how ComfyUI's Manager-installed packs work too, but you should know what you're clicking. The node's own GPU table covers Turing through Blackwell (RTX 20xx → 50xx, plus A100/A6000), so on a supported card "Install Nunchaku" is genuinely the fastest path to INT4 FLUX.

Installation

The installer node itself lives in ComfyUI-ArchAi3d-Qwen - ComfyUI Manager ("ArchAi3d Qwen") or:

cd ComfyUI/custom_nodes/
git clone https://github.com/amir84ferdos/ComfyUI-ArchAi3d-Qwen.git
cd ComfyUI-ArchAi3d-Qwen
pip install -r requirements.txt

Restart, find it under ArchAi3d/Utils. Personal use free, commercial license needed for paid work.

Where people get burned

Don't let "Check System" return a clean config and then assume the install is done - checking is free, installing changes your environment, and a failed mid-install can leave a half-installed package. If you're not on a supported GPU (or you're on one with the wrong PyTorch), the wheel install will fail loudly; the force_* overrides exist precisely for that. And if you're on a normal RTX 30/40 card with a working ComfyUI, ask yourself whether you need this at all - INT4 FLUX is compelling on 50-series and low-VRAM cards, but if your current setup already fits, an installer node isn't where you want to risk your working environment.