Dedicated GPU allocation
GPUtw rents a worker machine, or a single GPU on one. The GPU is exclusively yours either way — no other customer runs on the card you rent — and the container receives the CPU cores, usable RAM and instance storage that go with it, after host reserves.
Where a machine is shared, the deploy step says so on the card you pick: Dedicated GPU, shared system means the GPU is yours, the CPU, RAM and disk figures on that card are capped shares, and the host, its storage bandwidth and its network are shared with another tenant.
Storage
Container workspace storage is tied to the selected worker. The persistent Vault is mounted at /vault inside running GPU containers.
How much workspace space you actually have
/workspace is sized to the disk your instance rents. df inside the container reports something different — the machine's own disk — because that is the filesystem underneath. On a whole-machine rental nothing there belongs to another customer; on a shared one (Dedicated GPU, shared system) that disk is shared with another tenant. Either way the figure includes the platform's own usage (the OS, container images) and it is not the size your instance is held to.
Take the number your instance is actually measured against from its resources, which reports allocation and live usage side by side:
GET/api/instances/{id}/resourcesWriting past the rented size stops the instance rather than failing the write, so a job that trusts df can be interrupted while the disk still looks far from full. Anything you need to keep belongs in /vault, which has its own quota and survives the instance.
Shut down from inside the instance
Every instance built on a GPUtw base image carries gputw-shutdown, which shuts the instance down from a shell inside it. Chain it after a job and the GPU is released the moment the work finishes, instead of billing until someone remembers to stop it.
python train.py && gputw-shutdown- gputw-shutdown: stops the instance — the record is kept and /vault persists, so you can start it again later
- gputw-shutdown --terminate: deletes the instance outright
- gputw-shutdown --help: prints the same summary inside the container
/workspace is ephemeral either way — copy anything you want to keep into /vault before shutting down. Note that && only shuts down when the job exits 0; use ; if you want the instance to stop even when the job fails.
The command authenticates with an instance-scoped token the platform injects, so it can only ever shut down the instance it is running in — no API key, and no reach into anything else on your account. Custom images have the command only if they are built FROM a GPUtw base image; they still receive the token, so the same shutdown is one curl away (see Instance Runtime in the API section).
Ports
Templates can expose default web ports. You can add HTTP ports and raw TCP/UDP endpoints from the dashboard after launch.