Deploy from the App Gallery
Start from a curated app configuration and choose how to launch it.
The App Gallery provides starting configurations for supported workloads. A template supplies an app configuration that you can launch or customize for your own environment.
Choose a template
Select Select Template from Gallery, then choose a workload. Choose Custom Model comes first for an Ollama model of your choice. When no grid is available, templates that can run on serverless CPU are ordered first. The vLLM multi-GPU template category is currently hidden in both the gallery and templates API. Review the selected model, image and resource requirements.
Review the compute target
Select serverless CPU capacity or an available grid. GPU workloads need a grid; serverless GPUs are currently paused. Grid quotas count against the app’s creator. If serverless capacity is disabled for Developers, Quick Launch needs an accessible grid; ask an admin when none is available.
Launch or customize
Quick Launch submits the preconfigured app with the selected compute target. Use it when the settings already match your needs.
Customize First returns you to the configuration wizard. Review ports, environment variables, hardware and deployment location, then launch from Resources.
Try a small model on CPU
Supported small Ollama models, up to about 9B parameters, offer Try on CPU. For a GPU template without a GPU grid, the portal offers Start on CPU now when eligible, Set up a grid, or Not now. CPU replies can be slower than GPU replies. The organization must have credit and allow the member to use serverless; otherwise an admin needs to resolve the credit or compute restriction. A CPU trial changes the configuration to use no GPU.
When using list_app_templates, read notes before passing a template to create_app. GPU templates do not include a grid and need config.grids.ids; copying their configuration unchanged will not create a serverless GPU app.
Confirm that the app is ready
Open the app detail page. Follow its status, inspect startup logs and check the endpoint once the workload is ready. Model downloads and container startup can take time; an accepted launch does not mean that the service is immediately available.
If a template needs a token, password or other secret, configure it before sending production traffic. Do not put secrets in screenshots or support messages.
Last updated
Was this helpful?
