fast), cost per token at scale (throughput), or lowest cost (minimal). A deployment created from a shape starts from a configuration that is known to work for that model.
Why deployments created without a shape often fail
If you create a deployment without a shape — that is, without passing--deployment-shape (or deploymentShape in the API) — the configuration is not validated ahead of time. Mistakes only surface at creation, where they cause failures. Common examples:
- A GPU count that cannot fit the model in memory
- A context length the configuration can’t serve
- A quantization or precision the model doesn’t support on that hardware
How to find and use a shape
- firectl
- REST API
- Web UI
List the shapes available for your model:Then pass the shape’s name to
--deployment-shape when creating the deployment: