The ready-made environments — PyTorch with JupyterLab, Ollama, vLLM and others — cover the common cases. When you have your own stack, run any public Docker image instead, with a command that runs once, first, when the rental starts: the same idea as Vast's "onstart".

On the website

  1. Turn on Pro mode with the switch in the header.
  2. Choose a GPU on the marketplace and go to checkout.
  3. Open Your own Docker image and type the image: pytorch/pytorch:latest, nvidia/cuda:12.4.1-runtime-ubuntu22.04, ghcr.io/you/your-image:1.0…
  4. Optionally, add a start command under Run when it starts.
  5. Press Use this image, then pay.

A start command can install what your work needs and check the GPU, for example:

pip install transformers accelerate datasets && python -c "import torch; print(torch.cuda.get_device_name())"

Its output appears in the Console on the rental page. From there, work in the browser terminal or over SSH.

The image and start command are set at checkout and can only be changed before you pay; Use the environment above instead switches back to a ready-made environment.

What to expect inside

  • The image needs sh. Nearly every Linux image has it.
  • Your files live in /workspace, kept across commands for the whole rental and deleted when it ends — save what matters to your computer or to My Drive.
  • The image's own start-up does not run. The container starts with a shell and waits for your commands; its ENTRYPOINT and CMD are skipped. Put whatever must happen first into the start command.
  • The start command has up to an hour to finish, so use it for setup. Start a long-running server from the terminal or over SSH, with nohup … & so it keeps running after you disconnect.
  • Downloading the image is not billed. An image too big for the machine's free disk is refused with a message — choose a smaller one, or a machine with more free disk (every listing shows it).
  • Public images only: Docker Hub, GHCR or any registry that needs no login.
  • Privacy: the host can see which image you run, never your start command.

Choosing an image

  • Pin a version. pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime gives every rental the same software; latest can change under you.
  • Smaller starts sooner. The machine downloads the image first; that time is not billed, but you still wait for it.
  • Try it on your own computer first with docker run --rm -it IMAGE sh, to check that it starts and has what you need.
  • Reuse it for batches. The same image, passed as image, can run thousands of tasks over the pool.

Through the API

Start an instance with your image and start command. It is billed by the second until you stop it or duration_s runs out:

curl https://rentgpupower.com/api/v1/compute/instances \
  -H "Authorization: Bearer $GPURENT_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"image": "pytorch/pytorch:latest", "onstart": "pip install transformers accelerate", "duration_s": 3600}'

Or with the gpurent command line:

gpurent start --image pytorch/pytorch:latest --onstart "pip install transformers accelerate" --duration 3600
gpurent ssh 456
gpurent stop 456

gpurent start prints the instance's number (456 here), and gpurent instance 456 — or GET /instances/456 — shows its state and its cost so far. Stopping early returns the unused time to your credits, and filters such as --min-vram 24 choose the card. Your API key comes from the developers page; the full reference is in the docs.

FAQ

Can I use a private image?

No. Images must be public — on Docker Hub, GHCR or another registry that needs no login. Keep secrets out of the image.

Is the time spent downloading my image billed?

No. The time the machine spends downloading your image is not billed.

Why did my image's app not start by itself?

The rental starts the container with a shell, not with the image's ENTRYPOINT or CMD. Start the app yourself, from the terminal or over SSH — with nohup … & if it should keep running after you disconnect.

Does SSH work with my image?

Yes, if the image has a package manager (apt, apk, dnf or yum) to install the SSH server with, or already contains openssh-server — and the machine has the Docker sandbox.