Vast.ai

Vast.ai supplies training GPUs while your organization's connected storage keeps datasets and adapters. Tensorant searches for an offer and rents the instance when you launch a run. You do not need to rent an instance in Vast first.

An Owner can save connections in Tensorant. Open your name at the foot of the sidebar, then Settings → Connections under Organization. If you see Provided by Tensorant, contact your platform administrator about using your own connections.

1. Create and fund your account

1. Sign up or sign in to the Vast.ai console.
2. Open Billing from the account section and choose Add Credits. Complete payment and confirm a positive balance.
3. Budget for compute, disk storage, and transfers. Consider low-balance notifications or automatic top-ups. Unfunded instances and data can eventually be deleted; see Vast billing.

Creating a key and saving the Tensorant connection do not rent compute. Launching a training run is billable to this Vast account.

2. Create an API key

1. Open Vast's Keys page.
2. Under API Keys, choose +New and name the key, for example tensorant-training.
3. Scope it for offer discovery and instance management: searching offers, listing and reading instances, renting, starting, stopping, and destroying them.
4. Choose Create and save the key securely. See Vast's API-key guide.

The documented permission categories used here are misc for offer search, instance_read for discovery/status, and instance_write for launch and lifecycle operations. Vast's general instance-management example also includes user_read. Tensorant does not need billing transfers, key administration, or host-machine administration. For team accounts, ask the administrator to allow the required operations. See Vast permissions.

A read-only key cannot rent or release machines. Keep the key valid through cleanup and update Tensorant before revoking a key used by an active run.

3. Register SSH access for recovery

There is no SSH field in Tensorant's Vast form. Registering a public key in Vast gives you a way to recover local files if an upload fails.

1. Use an existing SSH key pair or generate one locally with ssh-keygen -t ed25519. Choose an unused file name if a key already exists and protect the private key with a passphrase.
2. Copy the public key from the .pub file. Keep the private key on your computer.
3. On Vast's Keys page, choose +New under SSH Keys, paste the public key, and save.
4. Confirm that it appears before launching training. Vast adds these keys to future instances; see SSH key setup.

4. Save the connection

1. In Tensorant, open Settings → Connections.
2. Choose Connect Vast.ai on its compute card.
3. Paste the key into Vast.ai API key, then choose Connect Vast.ai in the dialog.
4. Wait for Connected. Use Replace key to rotate an existing key.

This check reads account and marketplace information. It does not prove launch permissions or reserve an offer.

5. Connect permanent storage

Your organization's own storage credentials are supplied to the training machine for transfers. Their granted permissions apply there too. Allow for cross-cloud transfer charges.

The Vast local disk is working space. Tensorant downloads training data and exports adapters to the connected storage. This integration covers training; managed serving currently uses RunPod.

6. Verify a training trial

1. Create a project and dataset version using the quickstart.
2. Open Tune → Runs → Configure a run and choose Training trial · a small sample.
3. Select the dataset, set Training provider to Vast.ai, and choose an available Ampere-or-newer Training GPU that fits the model. T4 and other pre-Ampere GPUs cannot run the trainer.
4. Review Price limit ($/hour) and trial runtime, then Save trial draft. Saving does not rent compute.
5. When ready for a billable test, choose Launch trial, review the confirmation, and confirm.
6. Watch Training logs, then confirm adapter verification and instance cleanup.

Tensorant selects on-demand offers with one NVIDIA GPU, an amd64 host, CUDA 13 compatibility, sufficient disk and rental duration, a direct port, and at least 99% reliability. Only verified offers are considered. Tensorant supplies the image and startup command.

There is no location field in the current Vast connection form. Placement follows the matching marketplace offer rather than a required country or datacenter. If location must be fixed, use a provider connection with an explicit region.

Price and availability are rechecked at launch. A draft does not reserve an offer. Bandwidth can be charged separately from the displayed rental rate; see Vast pricing.

Recovery and billing

After a successful export, Tensorant verifies the adapter and destroys the instance. On a recoverable failure, it attempts to stop the instance while keeping its local disk. Stopped instances still incur storage charges. Destroying them removes local files; see Vast storage billing.

Use the Vast instance's connection or file-transfer instructions to retrieve adapter.zip. The Recover your adapter panel gives the container path: /workspace/data/adapter.zip with custom S3, or /workspace/data/runs/<run-id>/adapter.zip with RunPod storage. Upload it in Tensorant and choose Verify and recover. Restarting for recovery resumes GPU billing and depends on capacity.

If you do not need the output, use Cancel run and confirm cleanup. Wait for confirmed release. Revoking the API key or closing Tensorant does not destroy the instance.

Common problems