Models and endpoints
A deployment selects a saved provider, upstream model, public model name, transport and price. These are separate contracts.
Example
Contract
| Field | Role |
|---|---|
| model_name | Client model and routing/authorization name |
| litellm_params.model | Actual upstream model ID |
| litellm_credential_name | Saved provider connection |
| model_info.transport | Required registered executor |
| endpoint_types | Standard capabilities derive from transport; Bypass/Fal explicit |
Unknown/missing transports are rejected. Prices do not define capabilities. /public/endpoints is the registry; /model/available returns the current identity's actual executable endpoints.
Model discovery tries /models and /v1/models only after 404, within one 20-second limit. Other errors remain errors; an empty successful list is not retried. Discovery does not change the inference URL.
Source and verification
internal/provider/deployment/validate.go
go test ./internal/provider/... -count=1