Skip to main content

Model Files

Overview​

Model files are reusable, mountable, cacheable model data entries in the platform. They manage source, version, and mount relationships, and do not expose inference APIs themselves. When the same entry is reused across hosts, the platform can cache model data on those nodes for template pre-mounting and multi-deployment sharing, reducing repeated online downloads.

Relationships with templates, deployments, and images:

  • Inference Template: Mounts model files together with image and specs; Import Model often creates mountable model resources at the same time.
  • Inference Deployment: Starts the service from a template and loads mounted models; instances under a deployment are running replicas.
  • Inference Image: Must match the model and template type.

Console path: Artificial Intelligence → Inference → Model Files.

Common Operations​

Created via Inference Template Import​

Model files are created automatically when importing from Artificial Intelligence → Inference → Inference Templates → Import Model (model catalog / Hugging Face / ModelScope). The import creates mountable model resources and estimates required VRAM.

Save as Model File from an Instance​

Use this when a model has already been pulled and verified inside an instance. In the Inference Deployment sidebar, open the target deployment's Instances, go to the instance Models tab, and Save as model file (button label may vary). Fill in the name and save.

After saving, the entry can be selected in other same-type templates or deployments without pulling again.

View Details​

Sidebar tabs commonly include:

  • Details: Source, status, type, and mount/cache-related information.
  • Operation logs: Import, sync, and related task events.

Other Operations​

List or detail actions (subject to status and permissions):

  • Enable / Disable: Control whether the entry can be selected.
  • Set sharing / visibility: By project and visibility needs.
  • Enable auto-cache: Requires the entry to be enabled; uses extra disk on nodes.
  • Sync status: Pull the latest status from the platform.
  • Change project / Delete: If referenced by templates or deployments, unbind first; when changing versions, prefer adding a new entry then switching references.

FAQ​

Import is slow or fails​

Check network, bandwidth, proxy or firewall to the model source, model size, and node disk space. Online import requires nodes to reach the source; the platform environment may use a mirror. If outbound access is restricted, prepare the model on a networked instance first, then Save as model file. More troubleshooting: Quick Start / FAQ.

No selectable models in the template​

Confirm the list has ready entries of the matching type; models still preparing, or type mismatches with the template/image, cannot be selected.

Model not visible in the deployment after mount​

Confirm the mount took effect and the deployment / instance is ready; from the instance, call /v1/models (or the engine's model-list API) and check logs for mount or resource issues.

Delete failed or affects existing deployments​

Unbind model mounts from templates first, then delete. With multi-node or auto-cache enabled, watch disk usage.