Vast.ai setup

Marathon practise (production-scale training) runs on a GPU rented from Vast.ai. The studio talks to your Vast.ai account via an API key — it can see your rented instances and dispatch runs to them, but it can’t rent instances on your behalf.

One-time setup

1. Create a Vast.ai account

2. Get your API key

In the Vast.ai dashboard, go to Account → API Keys and copy your key.

3. Add the key to PFN Studio

In the Studio app, go to Settings → Vast.ai compute (/settings/vast) and paste the key. Hit Test connection to verify the studio can reach your account.

Renting a GPU

GPUs are rented separately — the API key just lets the studio see what’s on your account.

Pick a machine

Go to https://cloud.vast.ai/create/ and pick a GPU. For PFN training:

GPU	Typical hourly rate	Good for
RTX 3090	~$0.30 / hr	Most PFNs up to ~50M params
RTX 4090	~$0.50 / hr	Larger PFNs, faster training
A100	~$1.50 / hr	TCPFN-scale (200M+ params)

Rent it

Click Rent on the offer. The instance will appear in your dashboard within ~30 seconds and flip to running.

Verify in the studio

Back in the wizard’s Step 5 → Marathon → Pick a GPU, click ↻ Refresh. Your rented instance should appear, and you can pick it as the dispatch target.

After training

Marathon runs do not auto-terminate the instance. After your run finishes, head back to the Vast.ai dashboard and destroy the instance — otherwise you keep paying the hourly rate.

Troubleshooting

“No rented GPUs found” — API key is configured but no instances are visible. Either you haven’t rented one yet, or the instance is under a different Vast.ai account than the API key. Double-check at https://cloud.vast.ai/instances/.

“Couldn’t load Vast.ai instances” — the API key is invalid or rate-limited. Re-paste the key in /settings/vast and hit Test connection.

Run starts but never makes progress — the instance might be in a transitional state (booting, provisioning). Wait 30 seconds and refresh. If still stuck, destroy and rent a different one.