Yes it is possible. You can use minio, it's like s3/gs but it runs on a persistent volume of your on-premises storage.

Here are the instructions on how to use it as a kfserving inference storage:

Validate that minio is running in your kubeflow installation:

$ kubectl get svc -n kubeflow |grep minio
minio-service                                  ClusterIP   10.101.143.255   <none>        9000/TCP            81d

Enable a tunnel for your minio:

$ kubectl port-forward svc/minio-service -n kubeflow 9000:9000
Forwarding from 127.0.0.1:9000 -> 9000
Forwarding from [::1]:9000 -> 9000

Browse http://localhost:9000 to get to the minio UI and create a bucket/upload your model. Credentials minio/minio123. Alternatively you can use the mc command to do it from your terminal:

$ mc ls minio/models/flowers/0001/
[2020-03-26 13:16:57 CET]  1.7MiB saved_model.pb
[2020-04-25 13:37:09 CEST]      0B variables/

Create a secret&serviceaccount for the minio access, note that the s3-endpoint defines the path to the minio, keyid&acceskey are the credentials encoded in base64:

$ kubectl get secret mysecret -n homelab -o yaml
apiVersion: v1
data:
  awsAccessKeyID: bWluaW8=
  awsSecretAccessKey: bWluaW8xMjM=
kind: Secret
metadata:
  annotations:
    serving.kubeflow.org/s3-endpoint: minio-service.kubeflow:9000
    serving.kubeflow.org/s3-usehttps: "0"
  name: mysecret
  namespace: homelab

$ kubectl get serviceAccount -n homelab sa -o yaml
apiVersion: v1
kind: ServiceAccount
metadata:
  name: sa
  namespace: homelab
secrets:
- name: mysecret

Finally, create your inferenceservice as follows:

$ kubectl get inferenceservice tensorflow-flowers -n homelab -o yaml
apiVersion: serving.kubeflow.org/v1alpha2
kind: InferenceService
metadata:
  name: tensorflow-flowers
  namespace: homelab
spec:
  default:
    predictor:
      serviceAccountName: sa
      tensorflow:
        storageUri: s3://models/flowers