# ComfyUI-MultiGPU v2

A ComfyUI custom node that adds one-click virtual VRAM and multi-GPU distribution for UNet, CLIP and VAE loaders, offloading model layers to other devices to free VRAM for larger generations.

ComfyUI-MultiGPU adds custom nodes that let a UNet, CLIP or VAE loader offload part of a model to a donor device, such as system RAM or another GPU's VRAM, instead of keeping everything on the main compute card. This does not add parallel processing; workflow steps still run sequentially, but freeing VRAM this way lets a workflow handle a larger latent space, such as longer videos or more concurrent images, and can also speed things up by avoiding repeated model loading and unloading when VRAM is tight.

Its second version, called DisTorch2, adds native support for both plain safetensors models and GGUF-quantized models, up to ten percent faster GGUF inference than the first DisTorch approach, dedicated integration nodes for WanVideoWrapper, and two model-driven allocation modes, by byte count or by ratio, alongside the original virtual-VRAM slider.

The repository is set to be archived on 30 September 2026 and will no longer receive updates from its maintainer; fork maintainers are asked to announce themselves in a pinned coordination issue, though no fork is officially endorsed. It is written in Python and released under the GPL-3.0 license.

Website: https://github.com/pollockjj/ComfyUI-MultiGPU