Compute · cloud GPU

Runpod

Studio can create and terminate a Runpod GPU pod through the Runpod API. The pod pulls the same server image and exposes the Studio and API ports through Runpod's proxy hosts.

API deploy GPU account balance required

Referral link: it supports the project and does not change the price.

Before you start

A Runpod account with credit. Pod creation fails when the balance is too low, even if a card is on file.
The server image in a registry the pod can pull: Docker Hub is the simplest.
A Runpod API key (Settings → API Keys). Studio stores it encrypted.

"Your account balance is too low to rent a pod." Add credit in the Runpod console. Studio surfaces the API error; it cannot top up the account.

Push the image

The pod runs the combined server image. Publish it from Studio's Publish panel, or use Docker Hub directly:

docker tag sharpmind-server:gpu your-name/sharpmind-server:gpu
docker push your-name/sharpmind-server:gpu

The repository can stay private if the pod is given registry credentials. The simple path is a public repository. The tested reference image is integral2u/sharpmind-server:gpu.

Create the host

Settings → Providers → Compute → Add host → Runpod GPU.
Paste the Runpod API key. It is encrypted with the local key box.
Set the Server image to your-name/sharpmind-server:gpu.
Pick a GPU type (RTX 4090, L4, A40, A100 …) and a container disk size.
Save host.

The pod needs the network volume mounted at /data/SharpMind to keep models between restarts. Without it, the container disk is ephemeral.

Deploy the pod

On the host card press Deploy pod. Studio:

Creates a pod from the saved image, GPU type and disk.
Exposes HTTP ports 5117 (Studio) and 8080 (API).
Waits for the pod to reach RUNNING.
Fills the API and Studio URLs and runs Test.

The URLs look like https://<pod-id>-8080.proxy.runpod.net and https://<pod-id>-5117.proxy.runpod.net. The token in the container is the one saved with the host.

Use the pod

Upload a chat model from the card, or leave it empty for a fine-tune.
Chat or train with Run on = Runpod; sources and base models upload with the job.
Fine-tune exports download back to Studio; the pod does not need to stay alive afterwards.
Refresh reads the pod state; Logs is available through the remote Studio or Runpod's web terminal.

Terminate

Press Terminate on the card to stop billing. Runpod's proxy host is tied to the pod; a new deploy gets a new URL. Keep the network volume if you want the models to survive.

Network volume vs container disk. Models, uploads and training exports live under /data/SharpMind. Only a network volume survives pod replacement.

Problems

Message	Fix
Balance too low	Add credit in the Runpod console, then Deploy again.
Image pull fails	The image is private or the name/tag is wrong. Make it public or add registry credentials.
Pod stays CREATED / not RUNNING	No capacity for that GPU in the region, or the image cannot start. Try another GPU type.
API times out	Port 8080 was not exposed, or the pod is still starting. Check the pod's HTTP ports.
401 from the API	The host token and container token differ. Re-enter the token and redeploy.

Train on the pod → Compare providers