Verda¶

Verda is a GPU cloud with dedicated GPU and CPU instances (real VMs) in European data centers. Instances are billed hourly, either on demand or as spot capacity.

export VERDA_CLIENT_ID=... VERDA_CLIENT_SECRET=...   # or [verda] in config.toml
gpuprov provision --cloud verda --name dev --gpus H100:1
gpuprov provision --cloud verda --name cheap --gpus A100:1 --spot --region FI

Credentials are the Cloud API credentials (client ID and secret) from the Verda console's Credentials page. gpuprov exchanges them for a short-lived OAuth token and refreshes it automatically.

Planning: live availability¶

gpuprov reads the instance-type catalog and the current per-location availability (GET /v1/instance-availability, spot or on-demand to match --spot). Each location that has the type in stock becomes one offer:

instance_type = Verda's type, e.g. 1H100.80S.22V (1× H100 SXM5, 22 vCPUs)
region = the location code (FIN-01, FIN-02, ICE-01, …)
ordered by hourly price (the spot price with --spot), cheapest first

--region accepts location codes or ISO country codes (--region FI). No --region means anywhere. --gpus H100:8 picks types with exactly eight H100s; --instance-type takes an exact type, including CPU-only types.

Stock is re-checked for the chosen location right before deploying, so a stale offer fails without creating anything. A deploy that hits "no capacity" moves on to the next offer.

What gets created¶

Resource	Notes
instance	tagged `gpuprov-vm-name=<name>` and `gpuprov-managed=true`; hostname `gpuprov-<name>`
OS volume	`--disk` GB, tagged the same; deleted permanently with the instance
startup script	the k3s installer (see below), one per VM, removed on teardown
SSH key	your gpuprov public key, registered once and reused; never removed

The image is picked from the instance type's supported OS list: the newest Ubuntu with CUDA for GPU types, plain Ubuntu for CPU types. --image takes any Verda image type. You log in as root.

Kubernetes¶

Verda runs a startup script as root on first boot, so the k3s installer goes in as a startup script and k3s installs while the VM boots. gpuprov then verifies the install over SSH and falls back to pushing the installer itself if needed.

Verda has no firewall API, so gpuprov cannot guarantee that ports other than SSH are reachable. Clusters therefore use SSH-tunnel access (gpuprov tunnel <name>), and --port is rejected.

Teardown¶

gpuprov down deletes the instance together with its volumes (permanently, not to the trash). It then sweeps any volume still tagged for the VM, plus the VM's startup scripts. Teardown is retried until Verda confirms the instance is gone; otherwise the VM is recorded as LEAKED.