Operate a model

Use the model detail page as the source of truth. Its status and ready replica counts can change while provisioning, restarting, or applying an edit.

Read the lifecycle

Status	Meaning	What to do
Pending	The create request was accepted; no runtime is ready yet	Wait, or use Delete if the model is no longer needed
Scheduling	Runtime resources are being admitted and scheduled	Wait; inspect Logs if progress stops
Warming	Model assets are loading	Wait; startup duration depends on the model and configuration
Deploying	Serving resources or a new revision are progressing	Keep watching status, ready replicas, and Logs
Running	The route is published	Call the endpoint; check ready replicas if the page reports a degraded phase
Restarting or Stopping	A lifecycle action is still in progress	Do not issue another conflicting action
Stopped	The runtime is offline and requests are not served	Edit, Restart, or Delete
Failed	Provisioning or a lifecycle operation did not converge	Read the status reason and Deploy diagnostics, then use an action the Console offers
Finalizing	This model’s route update and runtime teardown are still in progress	Stop sending traffic to this model and wait. If another model serves the same Served Model Name, that route can remain callable
Finalized	This model’s route update and runtime teardown have completed	This model no longer serves requests. A same-name route can remain callable through another model; delete the remaining record when offered
Deleting	Permanent teardown is in progress	Wait for the list/detail page to confirm the resource is gone
Delete failed	Teardown could not finish	Retry Delete after checking the reported error

For client rollout, use the strict readiness gate: Running + endpoint present + complete Copy as cURL sample.

Scale or update configuration

Choose Edit on the detail page.

• For a single-node model, change Replicas to a whole number of at least 1.
• For a prefill/decode model, change the prefill and decode replica counts separately.
• Scaling adds or removes replicas in place; existing replicas are not restarted for a replica-only change.
• Runtime changes create a new revision and roll the model. Review the How to apply descriptions in the Console before saving; a single-replica model can be unavailable while it reloads.
• If the model is Stopped, edits are saved and take effect on the next Restart.

The current edit flow does not change the bound model, resource name, Served Model Name, or Flavor. Deploy another model when one of those values must change.

Do not change runtime tuning, image, environment, or engine flags by trial and error. Start from the model recipe and use a configuration validated for that model and Flavor.

Observe the model

Use the tabs for different questions:

• Copy as cURL — the current callable host, path, and Served Model Name.
• Config — the requested serving configuration. Runtime details can be restricted to catalog admins.
• Metrics — live pool health, ready/desired replicas, and the available metrics board.
• Logs — deploy diagnostics and serving output. Start here for Failed, a stalled rollout, or repeated 503 responses.

An empty panel is not proof of a healthy or unhealthy model. Correlate status, ready replicas, endpoint availability, request results, Metrics, and Logs.

The four tabs sit under the topology card. Logs is where a stalled rollout or a repeated 503 explains itself.

Stop, restart, or delete

Action	Use it when	Result
Stop	You want the model offline but may use it again	Requests stop; the model record and configuration remain
Restart	A model is Stopped, or the Console offers recovery from Failed	Runtime provisioning starts again; wait for the full readiness gate
Delete	You are finished with the model	This model is removed. The tenant endpoint and same-name route can remain for other models; the action cannot be undone

Delete can supersede a model still provisioning. Pending has no runtime to stop, so use Delete, not Stop, to cancel immediately after creation.

Cleanup checklist

When the model is no longer needed:

1. Save any diagnostics you are authorized to retain.
2. Choose Delete and confirm the irreversible action.
3. Wait until the model disappears from the list or the detail page reports that it no longer exists.
4. Review API Keys. An Inference key selected through this model authorizes its Served Model Name, so a replacement that continues using that name can still use it. Delete does not revoke the key; revoke it explicitly when that Served Model Name should no longer be authorized.
5. Remove the endpoint and key from local environment variables, application configuration, and secret stores that no longer need them.

The Can call column is the one to read at cleanup time: it names the scope a key still authorizes.

For failures during these operations, use Troubleshoot self-hosted models.