How to 'undeploy' a model from the Model garden?

[Louk]

After deploying a model through the model card in the Vertex AI Model Garden with the blue 'Enable' button, I now want do undeploy this model as it was only for testing purposes. However, a) no 'Unable' button appears on the model card of the deployed model, b) the "Publisher" model does not appear in either the "Vertex AI - deploy and use - model registry" or the "Vertex AI - Model Garden - View my Endpoints and Models" tab. Moreover, I'm unable to find a API endpoint to undeploy a "Publisher" model. A "Publisher" model is a model from e.g. Meta/Mistral/Anthropic hosted on GCP. How can I unable this model endpoint?

[dawnberdan]

Hi @Louk,

Welcome to Google Cloud Community!

You're correct that undeploying a "Publisher" model from Vertex AI Model Garden isn’t currently possible via the UI or a specific API endpoint.

Here’s an explanation why: 

• Model Garden Integration: "Publisher" models are connections to outside services like Meta, Mistral, or Anthropic. Instead of being directly set up in your Google Cloud Project, these models work through a proxy service that sends requests to the external provider.

• Limited Control: Because you lack direct control over the model deployment infrastructure, undeploying isn't a simple option.

For the workarounds, although you can’t undeploy directly, you can manage your usage by:

• Stopping Usage: The easiest approach is to cease sending requests to the endpoint. If you don’t need access to the model anymore, you can effectively "disable" it by not using it.
• Disabling Access Keys: If you're using access keys to authenticate with the external service (like API keys), consider revoking or disabling those keys to limit access.

For Future Consideration:

• Feature Request: Consider submitting a feature request to Google Cloud Platform (GCP) for the option to undeploy "Publisher" models. This could result in a dedicated API or UI feature being introduced.
• Alternative Models: If you want a model you can fully manage and undeploy, look into using models hosted within your GCP project or those from other sources that allow undeployment.

Additional Reminder:

• Even if you’re not actively using the "Publisher" model endpoint, you may still face charges (like those for the proxy service). Be sure to check your GCP billing to avoid any unexpected costs.
• If you're interested in exploring Model Garden, you can refer to this document.

I hope the above information is helpful.