Support ModelMetadata endpoint

[njhill]

From the KServe V2 Predict API:

Model Metadata:

GET v2/models/${MODEL_NAME}[/versions/${MODEL_VERSION}]

Details of the REST API are here and corresponding gRPC API here.

[rachitchauhan43]

@njhill : Do you have any docs regarding what exactly you are looking for in model metadata ? Or, maybe if you can put more details into this issue, that would be great.

I also had the same requirement.
For my use case what we are looking for is

1. How would a user know with what inputs the ISVC endpoint has to be called ?
2. How would user know what outputs ISVC endpoint would return ?

Few ideas I had were:

1. If a model is written using a serving runtime, then that particular serving runtime, should have implemented an endpoint like /get_metadata so that any model which is deployed using that serving runtime will always have model metadata endpoint that users can use to see i/p &o/p contracts.
   But this approach would fail when someone writes a custom predictor.
2. In case of a custom predictor, we have multiple options:

• Either we can mention the metadata as part of ISVC YAML and make it queryable like seldon does here, OR
• We tell users to create another endpoint inside their custom predictor as mentioned here in this thread in kserve cmty channel

Let me know your thoughts on this.

[njhill]

@rachitchauhan43 I'm just talking about extending the grpc-gateway in to cover the existing KServe V2 predict API Model Metadata endpoint in addition to passing through the Infer call:

GET v2/models/${MODEL_NAME}[/versions/${MODEL_VERSION}]

Details of the REST API are here and corresponding gRPC API here.

You're right that this assumes that the model runtime in question supports this API, for now custom runtimes would also need to implement.

I'll update the PR description to clarify a bit!