Source: https://docs.sogni.ai/run-a-worker/fast-worker/running-on-salad/

# Run a Sogni Comfy Worker on SaladCloud

Deploy the Sogni Docker image on Salad Container Engine using a **custom container group**. Start with your [Sogni worker key, Fast Worker NFT token ID, and hardware requirements](https://docs.sogni.ai/run-a-worker/fast-worker/running-sogni-fast-worker-remotely/#before-you-rent-a-gpu).

## [#](https://docs.sogni.ai/run-a-worker/fast-worker/running-on-salad/#create-a-container-group)Create a container group

1.  In the [SaladCloud Portal](https://portal.salad.com/), open your organization and project. Select **Deploy a Container Group** → **Custom**.
2.  Enter `sogni/comfy-worker-cu13:latest` as the image, or use the [recommended versioned image](https://api.sogni.ai/v1/worker-images/comfy?cuda=13). Leave the command override blank.
3.  Set **Replicas** to **1** and leave autoscaling off. Select a GPU class that meets your workload requirements. When you select several classes, Salad can assign any of them. Our RTX 5090 video configuration used **8 vCPUs, 60 GB of system RAM, and a 250 GB storage request**. Check the [requirements for your intended jobs](https://docs.sogni.ai/run-a-worker/fast-worker/hardware-requirements/) and allow space for the image and temporary downloads as well as the models.
4.  In **Environment Variables**, set `API_KEY` and `NFT_TOKEN_ID`. Set [workflow preferences](https://docs.sogni.ai/run-a-worker/fast-worker/sogni-fast-worker-advanced-configuration/#preferredworkflowids) before launch if you want to download only the models for selected jobs.
5.  Leave **Container Gateway** off. Under **Health Probes & Logging**, add the private HTTP probes below.
6.  Deploy the group and start it if autostart is off. Use the group’s logs to follow allocation and startup, then [verify the worker in Sogni](https://docs.sogni.ai/run-a-worker/fast-worker/running-sogni-fast-worker-remotely/#verify-the-deployment).

| Probe on port 8000 | Path | Initial delay | Period | Timeout | Successes | Failures |
| --- | --- | --- | --- | --- | --- | --- |
| Startup | `/startup` | 90 sec | 30 sec | 5 sec | 1 | 20 |
| Liveness | `/liveness` | 30 sec | 30 sec | 5 sec | 1 | 3 |

The startup check can pass while models are downloading. Use the Sogni dashboard to confirm the worker can accept jobs.

For the settings in the SaladCloud Portal, see [SaladCloud’s container group quickstart](https://docs.salad.com/container-engine/tutorials/quickstart).

## [#](https://docs.sogni.ai/run-a-worker/fast-worker/running-on-salad/#rescheduling-and-storage)Rescheduling and storage

Salad can move your container to a different machine, so treat its local storage as temporary. Expect the worker to download models again on a replacement node and keep the initial workflow selection focused. If the container repeatedly fails health checks, check the logs for driver, RAM, or storage errors before restarting it.

Create a separate group with a different NFT for another GPU. To end a rental, stop the container group in SaladCloud. For changes that restart the group, wait for the worker to finish its active job. See [Managing SaladCloud deployments](https://docs.salad.com/container-engine/how-to-guides/managing-deployments).
