ComfyUI-MultiGPU-Orchestrator
Smart multi-GPU routing for ComfyUI.
ComfyUI Multi-GPU Orchestrator
Use ALL the GPUs!
<p align="center"> <img width="48%" alt="ComfyUI-MultiGPU-Example 1" src="https://github.com/user-attachments/assets/c8bb0720-e7d3-4014-85b4-8f46461fb582" /> <img width="48%" alt="ComfyUI-MultiGPU-Example 2" src="https://github.com/user-attachments/assets/dad6aa42-3ce2-48a0-a2e2-c8a2b9a2d4d9" /> </p>Multi-GPU Orchestrator turns one ComfyUI session into a smart multi-GPU routing system. Perfect for cloud clusters (Vast.AI, Runpod) and CUDA-rich powerusers. Unlock maximum resource utilization by dispatching workflows across every GPU on your system. All you have to do is click the "► Run" button.
- One ComfyUI instance controls all GPUs.
- No workflow changes or added configuration required.
- Queue progress, history, and media assets are all available from the primary UI.
- Works well with cloud providers out of the box.
How It Works
When ComfyUI starts, the extension discovers visible CUDA devices and starts a ComfyUI worker process on each GPU. The main ComfyUI process stays as the browser-facing UI and orchestrator. Worker processes do the generation work.
The normal ComfyUI frontend is patched so queueing, status polling, job progress, and media assets are routed through the orchestrator automatically.
Direct API submissions work the same way: POST /prompt (and /api/prompt) on the main server is dispatched to the least-busy healthy worker. If no worker is available, the request falls back to the main ComfyUI process.
The Console keeps the standard Logs tab for the main process and adds a
GPU N tab for every worker. Worker output is written to
ComfyUI/logs/mgpu-workers/gpu-N.log; these files are cleared when the
orchestrator starts and reused for worker restarts during that session.
The MultiGPU sidebar settings can automatically respawn failed workers and, independently, re-queue only the jobs that were still running or pending in the worker's last queue snapshot. Completed jobs are removed from the replay ledger. Workers also stop when the primary ComfyUI server shuts down, restarts, or exits unexpectedly.
Install
Clone this repository into ComfyUI/custom_nodes/:
cd ComfyUI/custom_nodes
git clone https://github.com/obsxrver/ComfyUI-MultiGPU-Orchestrator.git