Run a Sogni Worker on Nosana

Nosana supplies the rented NVIDIA GPU. Sogni Comfy Worker connects it to the Supernet, receives eligible AI jobs, and credits rewards for completed work. You manage the rental in Nosana and the worker in the Sogni Workers dashboard.

As of September 2026, the Sogni Comfy Worker template is proposed in Nosana template PR #172 and is not yet available in the public catalog. Once it appears in Nosana Deploy, use the steps below. Its README also contains a complete setup guide. This template runs the Fast Worker service; it does not open a standalone ComfyUI editor.

#1. Prepare your two Sogni values

One GPU worker needs one unique NFT. Stop any other worker using that NFT before deploying here. An API key can be shared by workers on the same Sogni account; the NFT token ID must be different for each simultaneously running worker.

Keep your deployment confidential. Do not publish a configured job definition or upload it to public IPFS with your API key inside. Follow Nosana's confidential-job guidance, and never share credentials in screenshots or support messages.

#2. Choose the work you want to serve

Start with a focused model bundle. A rental worker can stay busy serving popular image models or MiniMax video; downloading the entire catalog adds startup time and storage cost. GPU capacity determines which jobs fit, but does not mean you should preload every model that fits.

Create or sign in to your account at Nosana Deploy and fund it for the rental. Select Create Deployment, choose Sogni Comfy Worker, and select a preset:

Krea images + Identity includes Krea 2 Turbo, Dark Beast Krea 2 v3, Krea 2 Identity Edit v1.2, and Dark Beast Krea 2 Identity Edit v1.2. Shared model files are downloaded once. This is a compact starting point for an image worker, including on a larger GPU.

Both MiniMax presets contain the same MiniMax H3 video and Music 3 model files. Their RAM checks differ because MiniMax Turbo, including FastH3, needs at least 47 GiB of system RAM on 24/32 GB GPUs; larger GPUs use the 30 GiB floor. See Which jobs your hardware can take for every model's VRAM and RAM requirements. System RAM and GPU VRAM are separate resources. A large GPU does not imply a proportionally large host-RAM requirement. Check the RAM actually available to the container, which can be less than the host's advertised total.

The sizes above are unique model-file totals from Sogni's catalog, checked September 14, 2026. They are not total disk requirements or measurements of an existing host's cache. Allow additional free space for the worker image, temporary downloads, generated files, and catalog changes. Size storage for the selected bundle and the space actually available on the assigned host.

All presets use an x86-64 Linux host and the CUDA 13 worker, requiring NVIDIA driver R580 or newer. Setup checks its preset's RAM floor, but does not enforce the driver or free disk space. Confirm the market's specifications and inspect the assigned host. A rejected host can cause a retry or delay; the setup check cannot reserve a suitable node.

#Optional: an even smaller bundle

The presets above are ready to use. To specialize further, change only the value following --comfyConfig in the resources operation before deploying:

FastH3-only includes its shared dependencies and omits other MiniMax tiers and music. Keep the MiniMax preset's RAM check for your GPU class. The single-image choices omit Identity Edit; use the image preset when you want all four image models. The worker advertises only workflows it can serve with its local files and hardware.

#3. Enter your values and deploy

1. Choose the GPU market and keep Replicas = 1.
2. Use the Infinite deployment strategy for an ongoing worker.
3. In the job definition form, find sogni-worker and expand Environment variables.
4. Replace REPLACE_WITH_YOUR_SOGNI_API_KEY in API_KEY with your Sogni API key.
5. Replace REPLACE_WITH_YOUR_FAST_WORKER_NFT_ID in NFT_TOKEN_ID with your numeric token ID.
6. Keep the chosen preset's setup and health settings, apart from an optional smaller bundle above. Review confidentiality and the rental cost, then deploy.

To run a second GPU, create a second one-replica deployment with another NFT. Raising the replica count copies the same NFT into every worker and makes them compete for one identity.

#4. Wait for setup and verify the worker

The template pins the Sogni-owned setup image, Nosana Worker Configurator 0.1.0, by its immutable digest. It checks available RAM, resolves the recommended Sogni worker image, and prepares the model resources. A cold host may need tens of minutes or longer to download them. Watch the progress in Nosana; repeated restarts can delay setup or move the job to another cold host.

Open Sogni Workers and select your NFT. Confirm its online status, reported version, GPU, and available models. Review Worker Health for startup progress and Job History as work arrives. Nosana showing a running container is not proof of a completed Sogni job.

The template's port 8001 is a dedicated health endpoint. /startup passes when startup is progressing normally, while /readiness passes when the worker can accept work. No public ComfyUI or worker control port is needed.

#Settings, updates, and storage

Open the worker's Settings tab in Sogni for supported operator controls. See Using the Workers Dashboard and Advanced Configuration. Workflow preferences choose among models already available to the worker. To change the preloaded bundle, change the Nosana preset or catalog selector for a replacement job. Keep the bundle focused on work you want to serve, and check Job History before adding more model families.

Each new Nosana job resolves Sogni's recommended, versioned worker image. A running job stays on its existing image until replaced. Read the Comfy Worker release notes before an intentional update and allow for queuing and another download after replacement.

Model files use Nosana's remote-resource cache under /data-models; reuse depends on the assigned host. This template does not add a portable persistent /data volume. Replacement can lose local worker state. Compatible dashboard settings have best-effort recovery through Sogni, but this is not a backup of the worker's entire data directory.

#If something goes wrong

Nosana rental charges are separate from Sogni rewards. Neither job volume nor cost recovery is guaranteed. Stop the deployment in Nosana when you want to stop renting the GPU; closing either dashboard does not stop the rental.

For help, see the Fast Worker FAQ, email [email protected], or join Sogni Discord.