Deploy custom model out of the scope of a Project/Space

In this description the term custom models will refer both to Model bring on watsonx.ai with BYOM feature and models created by the Tuning Studio.

Current situation

Custom models are in a catalog and it is possible to deploy them in a Project or a Space.

A promptTemplate that use a custom model cannot be tracked in AI Use case. Therefor we cannot use lifecycle and governance features on such a prompt template

With version 2.1 watsonx.governance will be able to track such promptTemplate in AI use cases. BUT, for such a prompt template to follow the usual governance lifecyle (dev --> validate --> operate) an instance of the custom model must be deployed, along side the Prompt Template in the corresponding project/space. As a consequence, an instance of the custom model is not shared among multiple PromptTemplate but dedicated to a single prompt template. --> This is not a scalable solution for an enterprise

Proposition

Deployment of custom models should be manage by an admin outside of the context of a project or a deployment space. That way, the model is available from any Project/Space, like it is the case for models shipped with watsonx. 

- In practice, an enterprise doesn't want to let the responsibility to deploy an LLM (and thus using GPUs) to any developer --> it must be managed centrally for the enterprise

- The value proposition of watsonx is to mutualise model deployment to optimise GPU utilisation. The same should apply to custom models.