endpoints/online/managed: how to deploy with multiple models / subfolders

[joaocc]

Which example? Describe the issue

We are trying to test migration from SDKv1 deployment to v2.

In the SDKv1 scenario we would use the SDK to download models to a local azureml-models, which would then be included in the model definition. Ex below:

./local/azureml-models
    /model-name-1
        /1
            some-somenet.onnx
        /2
            somenet.onnx
    /model-another-b
        /3
            another.onnx

This folder structure is the one created by azureml sdk when downloading models in preparation for a deployment.

$schema: https://azuremlschemas.azureedge.net/latest/managedOnlineDeployment.schema.json
name: blue
endpoint_name:  ep-name
model:
  path: ../../azureml-models
...

When trying to reproduce this with online endpoints, using for instance, cli/endpoints/online/managed, we receive the following message: Directory /.../.../azureml-models is empty. path or local_path must be a non-empty directory.

Going through the documentation here https://docs.microsoft.com/en-us/azure/machine-learning/how-to-deploy-managed-online-endpoints#use-more-than-one-model there is a mention to creating a model explicitly pointing to a folder with multiple models.

However, it seems that az ml model create -f model-bundle.yml produces a similar error

$schema: https://azuremlschemas.azureedge.net/latest/model.schema.json
name: model-bundle-1
path: ../../azureml-models

$ az ml model create -f model-bundle.yml --resource-group my_rg --subscription my_sub --workspace-name my_ws
Directory /.../.../azureml-models is empty. path or local_path must be a non-empty directory.

Any idea how CLIv2 can be used to address this scenario? Thanks