Run a Sogni Comfy Worker on Nebius
Nebius AI Cloud gives you a Linux GPU VM with SSH access. Use the standard Sogni Linux installer on the VM. It selects the worker image, creates the Docker Compose deployment, and sets up the model cache. Prepare your Sogni worker key and Fast Worker NFT token ID.
#Create a GPU virtual machine
- In the Nebius console, go to Compute β Virtual machines and create a VM. Select With GPUs and pick a preset with one GPU. Weβve run the worker on Nebius RTX PRO 6000 and B300 instances. Choose a region and preset that meet your VRAM, RAM, and bandwidth requirements.
- For the boot disk, choose Ubuntu 24.04 LTS for NVIDIA GPUs (CUDA 13) where available for your preset. Size the disk for Docker, the workerβs files, and the selected models. To keep the worker files and model cache when replacing the VM, attach and mount a separate persistent volume and install the worker under that mount.
- Configure outbound Internet access and a way to reach the VM over SSH. Add your SSH public key under Configuration β Access. If you use a public IP, restrict inbound SSH to trusted addresses and leave the worker ports closed.
- Review your price and create the VM. Choose Regular for ongoing service. A preemptible VM can stop while the worker is serving a job.
See Nebiusβs single-GPU VM walkthrough and GPU boot disk images for the current console steps.
#Install and check the worker
- Connect over SSH. Run
nvidia-smiand confirm the expected GPU and R580 or newer driver. Check that Docker can see the GPU using the test in the Sogni Linux installation guide. Install and configure Docker Engine or NVIDIA Container Toolkit as described there if either is missing. - Follow the Linux guide to extract the installer on the disk you sized for the worker. Run
./worker-install.sh, select Comfy Worker, and enter the worker key and NFT token ID when prompted. - Verify the worker in Sogni. If it stays offline, run
./worker-debug.shfrom the installation directory and check GPU access, free disk space, and the identity settings.
When youβre done, wait for the active job to finish and stop the VM in the Nebius console. Check for retained disk and other resource charges, and confirm how your boot disk is managed before deleting the VM.