Is your feature request related to a problem? If so, please describe.
As part of model registry initiative[1] we started exploring[2] the new ClusterStorageContainer capability introduced since KServe 0.11[3].
I was able to make it working for plain Kserve deployments but I am struggling making it working using ModelMesh deployments as it seems custom URIs are not supported; I keep getting failed to fetch CR from kubebuilder cache for predictor fraud: the InferenceService test-model-registry-initializer/fraud has an unsupported storageUri scheme model-registry while trying to provide a custom URI like model-registry://....
If the feature is not supported with ModelMesh yet, I would suggest to extend that feature here too such that the ClusterStorageContainer capability works for both deployment modes.
Is your feature request related to a problem? If so, please describe.
As part of model registry initiative[1] we started exploring[2] the new
ClusterStorageContainercapability introduced sinceKServe 0.11[3].I was able to make it working for plain Kserve deployments but I am struggling making it working using
ModelMeshdeployments as it seems custom URIs are not supported; I keep gettingfailed to fetch CR from kubebuilder cache for predictor fraud: the InferenceService test-model-registry-initializer/fraud has an unsupported storageUri scheme model-registrywhile trying to provide a custom URI likemodel-registry://....After some inspection in the source code I think that the URI parsing pass through this method https://github.com/kserve/modelmesh-serving/blob/6c86da9473d50de63f9ea3af8a4d7c223849547e/pkg/predictor_source/inferenceservice_registry.go#L131 and I couldn't find any way to workaround that, therefore I started thinking that the
ClusterStorageContainerfeature is not supported withModelMeshdeployments.Is this true?
Is there any fix or trick, that I might have missed, to make it working?
[1] https://github.com/kubeflow/kubeflow/issues/7396 [2] https://github.com/kserve/kserve/issues/3343 [3] https://kserve.github.io/website/master/modelserving/storage/storagecontainers/
Describe your proposed solution
If the feature is not supported with
ModelMeshyet, I would suggest to extend that feature here too such that theClusterStorageContainercapability works for both deployment modes.Describe alternatives you have considered
n/a
Additional context
Custom model registry storage container example: https://github.com/lampajr/model-registry-storage-initializer