ComfyUI-MultiGPU

Why we included this project

This set of ComfyUI custom nodes helps when your GPU runs out of VRAM on larger image latents or longer video generations. Instead of loading the full UNet, CLIP, and VAE on your main compute card, it moves static model components onto system RAM or a second GPU, so the freed memory goes to actual computation. A single virtual-VRAM slider handles the offload automatically for casual use, while expert modes let you assign exact byte or ratio allocations across devices. It supports both standard .safetensors checkpoints and GGUF-quantized GGML models, and it ships dedicated multi-GPU nodes for WanVideoWrapper pipelines. Dual-GPU owners and people pushing high-resolution video work get the most out of it, but the CPU-RAM donor path also helps anyone stuck on a single card with limited video memory.