Templates

A template is the software environment your instance boots into. You pick one in step 2 of the deploy wizard, and it decides what you get when the machine comes up: a browser UI, a notebook server, or a bare shell with CUDA.

A template is a container image plus an optional start command. When you launch an instance, Grafilab runs that image on the node you picked and hands the container the GPU.

Three things are true of every template, because the platform adds them rather than the image:

• You are root. The container runs as root with a TTY attached, so you can install packages, edit config, and run anything the image lacks.
• SSH is always available. The platform layer puts an SSH server inside the container with a generated password, so even a template with no web UI is reachable. See Manage Instances for the Connect dialog.
• The GPU is passed through. nvidia-smi works inside the container on a GPU node, including in templates you would not think of as GPU tools, such as code-server.

Templates that ship a web UI declare a service on port 80, which Grafilab’s proxy publishes at https://<your-slug>.instance.grafilab.ai — the slug is the three-word name on your instance card. That mapping is part of the template. Any other port you start yourself needs a Web Access mapping, covered in Manage Instances.

Template	Image	Interface	Pick it when
Open WebUI + Ollama	grafilab/openwebui:0.10.2	Web UI + SSH	You want a ChatGPT-style chat window running open-source models privately on your own GPU.
JupyterLab	grafilab/jupyterlab:2026-06-09	Web UI + SSH	You work in notebooks — exploration, fine-tuning experiments, data work with CUDA and PyTorch ready.
ComfyUI	grafilab/comfyui:2026-06-09	Web UI + SSH	You generate images or video with node-based Stable Diffusion / Flux workflows. Bring your own models.
code-server	grafilab/code-server:4.123.0	Web UI + SSH	You want VS Code in the browser with an integrated terminal on the GPU box.
PyTorch + CUDA	pytorch/pytorch:* (four tags)	SSH only	You run scripts and training jobs and want a clean, unopinionated PyTorch image.

Two legacy llama:3.2.3b and llama:3.3 llamafile templates remain in the list for existing users. They are not maintained and are not documented here — use Open WebUI + Ollama instead.

The catalogue is served live, so the wizard is the authority on what is available today. Tags are pinned to a version or build date rather than latest, so a template behaves the same next month.

If none of the built-ins fit, bring your own image: in step 2 of the wizard click + → Create New Template, or manage the full list on the Docker Templates page. Your saved templates appear in the grid with a My Template badge. The image must run as root, must start on its own, and should carry a pinned tag; private images need Registry Credentials.

Custom Templates walks through the form, the credentials sub-modal, and the management page.

• A root shell over SSH, with the image’s own environment variables loaded into interactive bash sessions. Non-bash shells such as busybox ash do not pick them up.
• nvidia-smi on GPU nodes, for confirming the card matches what you rented.
• 2 GB of /dev/shm, which matters for PyTorch DataLoader workers. If you hit shared-memory errors, lower num_workers.
• No persistent volume.