Connections

Connections are the outside accounts your organization works with. GPUs start in your RunPod account and are billed there, and your data is kept on your RunPod network volume. Open Settings at the foot of the sidebar and choose Connections.

Connect compute first, then storage. Endpoints and the Hugging Face token can be added any time. A new organization shows a Set up checklist that walks you through these steps.

Before you start

• The owner role. Only owners set up, replace and remove connections. Members can test them. See roles.
• A RunPod account with a payment method.
• The right datacenter. Every GPU your organization starts runs in your volume's datacenter, and that datacenter must offer RunPod's S3 API. Pick one that has the GPUs you want. RunPod lists them in its S3 API documentation.

Own connections and shared connections

Most organizations bring their own connections, as this page describes. Tensorant can instead set up an organization to use connections we provide. Its Connections screen then says the platform administrator manages them: there's nothing to set up, you can still test storage and endpoints, and publishing adapters to Hugging Face isn't available.

Compute

Connect RunPod

1. In RunPod, open Settings → API Keys and create a key with read and write access.
2. In Connections, on the Compute card, choose Set up compute.
3. Paste the key into RunPod API key and choose Connect RunPod.

The key is tested with RunPod before it's saved. The card then shows the key's last four characters and your organization's limits: how many training runs and deployments can run at once, and for how many hours.

Test compute checks the saved key again, for example if you think it was deleted in RunPod.

RunPod balance

The Compute card shows how much credit is left in your RunPod account, and what it spends per hour when RunPod reports it. Refresh reads it again. Everyone in the organization can see it.

When the balance runs out, RunPod stops your training runs, deployments and published models. To warn you in time:

• Below $1, the card turns amber and every page shows a warning with a link to add funds in RunPod.
• Tensorant checks the balance about every 15 minutes. The first time it's below $1, the organization's owners get an email. They get another only after the balance went back to $2 or more and then fell below $1 again.

Add funds on RunPod's billing page. The warning goes away once the balance is $1 or more.

What RunPod billed this month

Two lines show what RunPod billed your account this month, outside training runs and deployments:

• Serverless this month, on the Compute card: the Serverless endpoints of your wake-on-request published models.
• Storage this month, on the Data storage card: your network volume.

Storage is kept separate because every run and deployment shares the volume, so no run is charged for it. Training runs and deployments show their own spend. See Estimated and actual spend.

Tensorant reads RunPod's billing about every 30 minutes. RunPod's billing can take a while to update, so each line says when it was last synced, and "may still change" until the amount has settled. Before the first sync, it says "not synced from RunPod yet".

Replace or remove the key

• Replace key: while anything is running, the new key must belong to the same RunPod account, so running GPUs can still be stopped.
• Remove compute: only possible when no training run, deployment or published model is running. Data on your volume is not touched.

Data storage

Create the volume and its S3 key in RunPod

1. Create a network volume in a datacenter that offers the S3 API and the GPUs you want. Note its ID and datacenter, for example EU-RO-1.
2. In RunPod's settings, create an S3 API key. It's separate from your API key and has an access key and a secret.

Connect storage

1. On the Data storage card, choose Set up storage.
2. Enter the Network volume ID, the Datacenter as RunPod shows it, the S3 access key and the S3 secret key.
3. Choose Connect storage.

Tensorant checks that the volume exists in that datacenter in your RunPod account and that the S3 keys can reach it. Once saved, your organization can create projects.

Test storage connection

Test storage connection repeats both checks. It doesn't start a GPU.

Replace keys or remove storage

• Replace keys saves new S3 keys at any time. Once you have projects, the volume and datacenter can't change, because your data is on that volume.
• Remove storage is only possible while the organization has no projects.

Removing storage doesn't delete files on your volume. RunPod keeps billing the volume until you delete it in RunPod.

Model endpoints

A model endpoint is any service that answers OpenAI-compatible chat completions: a model provider, or a model you run yourself. Recipes use one to generate examples, evaluations can use one as a judge or as the model being tested, and the Playground can talk to one. The provider receives the text you send it and bills you for it.

Add an endpoint

1. Under Model endpoints, choose Add endpoint.
2. Enter a Name, such as generator. Recipes and evaluations refer to it by this name, so it can't be changed later. A few names are reserved, such as compute, storage and playground, and names can't start with managed:.
3. Enter the Base address, such as https://api.example.com/v1.
4. Enter the Model name the endpoint expects.
5. Enter the API key, or leave it empty if the endpoint needs none.
6. Choose Add endpoint.

New endpoints show Untested until you check them.

Address rules

• Use https:// on port 443, with no user name, password, query or fragment.
• Use a public host name, not an IP address or a local name.
• If you change an endpoint's host, enter its API key again. A saved key is never sent to a different host.

Check connection

Check connection sends one short chat request and waits up to 20 seconds. The provider bills it like any other request.

Your project's ready deployments also appear in this list, marked Managed. See Deployments.

Edit or remove an endpoint

• Edit changes the address, model or key. Leave API key empty to keep the saved one. Evaluations and recipes that use the endpoint keep working only while its address and model stay the same.
• Remove deletes it. Anything that uses it fails until an endpoint with the same name, address and model exists again.

Hugging Face token

1. On huggingface.co, open Settings → Access Tokens and create a token. A read token is enough for importing; publishing adapters needs write access.
2. On the Hugging Face card, choose Add token, paste it and choose Add token.

The card then shows the Hugging Face account it belongs to. The token is used to import private and gated datasets and to publish adapters. Public datasets don't need it, and base models must be public.

Status and secrets

Keys, secrets and tokens are stored encrypted and never shown again, not even to owners. The console shows only their last four characters, who set them and when.

Limits

When something goes wrong