Templates

GPUtw templates define the container image, default Web UI behavior, exposed UI port, and the starting software stack. Common templates include PyTorch/Jupyter, TensorFlow, ComfyUI, Ollama/Open WebUI, CUDA development, and Ubuntu base images.

Templates with a browser app expose their configured Web UI port, commonly 8080. Use the dashboard Web UI button to open the authenticated URL instead of raw-exposing the same port.

Base images such as Ubuntu or CUDA development images may start with no web server. Add HTTP ports only when your process is listening inside the container.

You can bring a public or private registry image from Docker Hub, GHCR, or another container registry. Use an explicit non-latest tag or a sha256 digest. Custom image pull time is billed from deployment start, private registry credentials are removed after the instance lifecycle, and the image cache is best-effort removed from the worker when the instance stops.

• SSH: with SSH enabled you log in as root with your account SSH keys — they are injected automatically, and if the image has no SSH server one is installed at startup (the image must run as root with a package manager such as apk/apt/dnf)
• Environment variables: set KEY=VALUE pairs in the deploy form; they are injected into the container
• Startup args: override the image CMD one argument per line; the ENTRYPOINT is preserved
• Web UI port: must be 1024–65535 — an image serving on a low port (e.g. nginx on 80) needs its args or config changed to listen on a high port

If a framework, CUDA version, package set, or launcher is missing, use the template request flow from the templates page or contact support with the image requirements.