Search the Vast.ai marketplace for GPU offers, rent an instance, connect, and destroy it

Documented steps

1. Install the CLI: pip install vastai. Authenticate with vastai set api-key <API_KEY>, or send 'Authorization: Bearer <VAST_API_KEY>' for direct HTTP calls.
2. Search offers from the CLI: vastai search offers 'gpu_name=RTX_4090 num_gpus=1 verified=true rentable=true' -o 'dlperf_usd-' to sort by performance per dollar.
3. Or search over HTTP: POST https://console.vast.ai/api/v0/bundles/ with filter fields such as verified, rentable, num_gpus, gpu_name, type (ondemand, bid, or reserved) and dph_total.
4. The response contains an offers array. Each offer's id is the ASK id used to rent - it is not the instance id you will manage afterward.
5. Rent the offer: PUT https://console.vast.ai/api/v0/asks/{id} with a body whose only required field is image (the Docker image). Common optional fields: disk (GB), runtype (default ssh; also jupyter, args, ssh_proxy, ssh_direct), label, env, onstart, args_str, template_hash_id and volume_info.
6. Equivalent CLI form: vastai create instance <OFFER_ID> --image pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime --disk 20 --ssh --direct --onstart-cmd "nvidia-smi".
7. A successful rent returns HTTP 200 with new_contract - that value is the instance id for all later calls.
8. Poll GET https://console.vast.ai/api/v0/instances/{INSTANCE_ID}/ (or vastai show instance <id>) until the status reads running.
9. Get connection details with vastai ssh-url <INSTANCE_ID>, or read ssh_host and ssh_port from the instance object, then ssh root@<SSH_HOST> -p <SSH_PORT>.
10. Destroy when finished: DELETE https://console.vast.ai/api/v0/instances/{INSTANCE_ID}/ or vastai destroy instance <INSTANCE_ID>. This is what actually ends all billing.
11. Official documentation: https://docs.vast.ai/api-reference/instances/create-instance | https://docs.vast.ai/api-reference/search/search-offers | https://docs.vast.ai/cli/hello-world