Metadata-Version: 2.1
Name: comfyui-nunchaku-rocm
Version: 1.0.0
Summary: Run Nunchaku 4-bit SVDQuant models on AMD Radeon and Instinct GPUs (ROCm): up to 3.6x faster than stock ComfyUI on a 16 GB RX 7800 XT. Supported models: FLUX.2 [klein] 4B and 9B, FLUX.1-schnell, dev, Krea and Kontext, Qwen-Image, Qwen-Image Lightning and Qwen-Image-Edit, Z-Image-Turbo, SDXL and SDXL-Turbo. Includes a LoRA loader for all of them, including speed LoRAs (LCM, Qwen-Image Lightning). Setup: install here and restart ComfyUI (Manager adds nunchaku-rocm, plus Triton on Windows, without touching your PyTorch). On Windows its install.py also adds the ROCm SDK headers and Python files Triton needs. Then open Templates > Extensions > comfyui-nunchaku-rocm: each example workflow lists its model downloads. The first image of each model compiles kernels for a minute or two.
Home-page: https://github.com/TwelveNights/ComfyUI-nunchaku-rocm
Requires-Python: >=3.10
