Integrating MMAR and ngc

[vikashg]

Point of contact for NGC: Kristina Joos
It is possible to load the models and it works as a drive, where you can upload anything you want. So, it is really flexible for our needs. However, we need to define a good set of guidelines and what metadata information can be shared to run the model.
The only way to share the model publicly is by using the wget command. So for the moment maybe we can share the wget links for corresponding models.
There are some questions about how to curate the models so that the repository does not become populated with "junk" models. How to maintain a good repo practices.
Should we use some automated tools like the one used in MLFlow, Hugging Face.
How to maintain version history of the model and its dependencies.
What are the minimum requirements to run a model?
Should they be containerized ?

These are some of the points that came in Dec 10th meeting.