Transcript of Use ALL Your GPUs: ComfyUI Distributed Tutorial

Hello and welcome. Today I'm delighted to introduce you to Comfy UI Distributed, an extension that addresses a common frustration, having multiple GPUs yet being limited to using just one at a time. We'll explore how to harness every GPU at your disposal, whether they're in the same machine or distributed across your network. Let's learn how it works. First, let's install the extension. Navigate to your Comfy UI custom nodes directory and clone the repository. Upon restarting Comfy UI, you'll notice a new panel has gracefully appeared in your sidebar, the distributed GPU panel. Adding a local worker is refreshingly straightforward. In the distributed GPU panel, click add worker. Give your worker a descriptive name. Perhaps studio GPU 2. Assign it a unique port. 8189 works well. Then specify the CUDA device index which you can discover using Nvidia SMI. When launching your main Comfy UI instance, if you haven't specified a CUDA device, it will use CUDA device zero. So if you have two cards, set the local worker to be one. Save your configuration and with a single click, launch your worker. The interface provides real-time feedback with status indicators that pulse gently when workers are starting. This visual language helps you understand your system state at a glance. Green for running, red for stopped, and orange during processing. Now, let's integrate distributed processing into your workflow. The beauty lies in its simplicity. Add a distributed seed node and connect it to your sampler. This ensures each worker generates unique variations. Then place a distributed collector after your VAE decode. These two nodes are all you need to transform any workflow into a distributed powerhouse. For those working with highresolution imagery, the ultimate SD upscale distributed node offers something truly special. It intelligently divides your image into tiles, distributing them across available GPUs. The result, much faster upscaling. Remote workers open even more possibilities on your remote machine. Launch Comfy UI with the listen and enable cause header flags. This is essential for network communication. Configure your firewall to allow the worker ports. Then simply add the remote worker on your main machine using its IP address. The system handles the rest. Optionally, if the remote machine has multiple GPUs as well, add them as workers, ensuring the listen flag is used and ports are open through the firewall. Finally, add the remote worker's second GPU to your main machine. Throughout your distributed journey, the UI provides thoughtful touches. Clear memory across all workers with a single button. monitor individual worker logs in real time. Should you encounter any challenges, the built-in debug mode provides comprehensive logging. Enable it through the settings panel, and the system will guide you through any configuration nuances. You'll find the GitHub repository and comprehensive documentation in the description below. I warmly invite you to share your experiences and any questions in the comments. Your insights help shape this tool's evolution. Thank you for joining me today. May your creative explorations be swift, your iterations plentiful, and your GPUs active. Until next time, happy creating.