doc: Add custom runtime with PVC example (#495)

#### Motivation

I'm very happy to quickly install `kserve modelmesh` by following
[quickstart](https://github.com/kserve/modelmesh-serving/blob/main/docs/quickstart.md),
but I encountered the problem when I want to write a python-based custom
`ServingRuntime` for model `mnist-svm.joblib` which is also used in many
guides and docs, and I also opened an [issue
](#494) for it.
However, I can hardly find the complete process which is available
online, including the community. So I pieced together some information
to make it's easier for user to do it.

#### Modifications

1. Add a REAMD for describing the complete process
2. Add a folder for custom ServingRuntime, including the python code,
Dockerfile and required library configuration file.

#### Result

The process for writing a python-based custom `ServingRuntime` is
completed, hope it can be a easy start for fresh user.

---------

Signed-off-by: zhlsunshine <[email protected]>

docs/examples/python-custom-runtime/Python-Based-Custom-Runtime-with-Model-Stored-on-PVC.md

@@ -0,0 +1,279 @@
+# The Python-Based Custom Runtime with Model Stored on Persistent Volume Claim
+
+This document provides step-by-step instructions to demonstrate how to write a custom Python-based `ServingRuntime` inheriting from [MLServer's MLModel class](https://github.com/SeldonIO/MLServer/blob/master/mlserver/model.py) and deploy a model stored on persistent volume claims with it.
+
+This example assumes that ModelMesh Serving was deployed using the [quickstart guide](https://github.com/kserve/modelmesh-serving/blob/main/docs/quickstart.md).
+
+# Deploy a model stored on a Persistent Volume Claim
+
+Let's use namespace `modelmesh-serving` here:
+
+```shell
+kubectl config set-context --current --namespace=modelmesh-serving
+```
+
+## 1. Create PV and PVC for storing model file
+
+```shell
+kubectl apply -f - <<EOF
+---
+apiVersion: v1
+kind: PersistentVolume
+metadata:
+  name: my-models-pv
+spec:
+  capacity:
+    storage: 1Gi
+  accessModes:
+    - ReadWriteMany
+  persistentVolumeReclaimPolicy: Retain
+  storageClassName: ""
+  hostPath:
+    path: "/mnt/models"
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: "my-models-pvc"
+spec:
+  accessModes:
+    - ReadWriteMany
+  resources:
+    requests:
+      storage: 1Gi
+EOF
+```
+
+## 2. Create a pod to access the PVC
+
+```shell
+kubectl apply -f - <<EOF
+---
+apiVersion: v1
+kind: Pod
+metadata:
+  name: "pvc-access"
+spec:
+  containers:
+    - name: main
+      image: ubuntu
+      command: ["/bin/sh", "-ec", "sleep 10000"]
+      volumeMounts:
+        - name: "my-pvc"
+          mountPath: "/mnt/models"
+  volumes:
+    - name: "my-pvc"
+      persistentVolumeClaim:
+        claimName: "my-models-pvc"
+EOF
+```
+
+## 3. Store the model on this persistent volume
+
+The sample model file we used in this doc is `sklearn/mnist-svm.joblib`.
+
+```shell
+curl -sOL https://github.com/kserve/modelmesh-minio-examples/raw/main/sklearn/mnist-svm.joblib
+```
+
+Copy this model file to the `pvc-access` pod:
+
+```shell
+kubectl cp mnist-svm.joblib pvc-access:/mnt/models/
+```
+
+Verify the model exists on the persistent volume：
+
+```shell
+kubectl exec -it pvc-access -- ls -alr /mnt/models/
+
+# total 348
+# -rw-rw-r-- 1 1000 1000 344817 Mar 19 08:37 mnist-svm.joblib
+# drwxr-xr-x 1 root root   4096 Mar 19 08:34 ..
+# drwxr-xr-x 2 root root   4096 Mar 19 08:37 .
+```
+
+## 4. Configure ModelMesh Serving to use the persistent volume claim
+
+Create the `model-serving-config` ConfigMap with the setting allowAnyPVC: true:
+
+```shell
+kubectl apply -f - <<EOF
+---
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: model-serving-config
+data:
+  config.yaml: |
+    allowAnyPVC: true
+EOF
+```
+
+Verify the configuration setting:
+
+```shell
+kubectl get cm "model-serving-config" -o jsonpath="{.data['config\.yaml']}"
+```
+
+# Implement the Python-based Custom Runtime on MLServer
+
+All of the necessary resources are contained in [custom-model](./custom-model), including the model code, and the Dockerfile.
+
+## 1. Implement the API of MLModel
+
+Both `load` and `predict` must be implemented to support this custom `ServingRuntime`. The code file [custom_model.py](./custom-model/custom_model.py) provides a simplified implementation of `CustomMLModel` for model `mnist-svm.joblib`. You can read more about it [here](https://github.com/kserve/modelmesh-serving/blob/main/docs/runtimes/mlserver_custom.md).
+
+## 2. Build the custom ServingRuntime image
+
+You can use [`mlserver`](https://mlserver.readthedocs.io/en/stable/examples/custom/README.html#building-a-custom-image) or `docker` to help to build the custom `ServingRuntime` image, and the latter is done in [Dockerfile](./custom-model/Dockerfile).
+
+To build the image, execute the following command from within the [custom-model](./custom-model) directory.
+
+```shell
+docker build -t <DOCKER-HUB-ORG>/custom-model-server:0.1 .
+
+```
+
+> **Note**: Please use the `--build-arg` to add the http proxy if there is proxy in user's environment, such as:
+
+```shell
+docker build --build-arg HTTP_PROXY=http://<DOMAIN-OR-IP>:PORT --build-arg HTTPS_PROXY=http://<DOMAIN-OR-IP>:PORT -t <DOCKER-HUB-ORG>/custom-model-server:0.1 .
+```
+
+## 3. Define and Apply Custom ServingRuntime
+
+Below, you will create a ServingRuntime using the image built above. You can learn more about the custom `ServingRuntime` template [here](https://github.com/kserve/modelmesh-serving/blob/main/docs/runtimes/mlserver_custom.md#custom-servingruntime-template).
+
+```shell
+kubectl apply -f - <<EOF
+---
+apiVersion: serving.kserve.io/v1alpha1
+kind: ServingRuntime
+metadata:
+  name: my-custom-model-0.x
+spec:
+  supportedModelFormats:
+    - name: custom_model
+      version: "1"
+      autoSelect: true
+  multiModel: true
+  grpcDataEndpoint: port:8001
+  grpcEndpoint: port:8085
+  containers:
+    - name: mlserver
+      image: <DOCKER-HUB-ORG>/custom-model-server:0.1
+      env:
+        - name: MLSERVER_MODELS_DIR
+          value: "/models/_mlserver_models/"
+        - name: MLSERVER_GRPC_PORT
+          value: "8001"
+        - name: MLSERVER_HTTP_PORT
+          value: "8002"
+        - name: MLSERVER_LOAD_MODELS_AT_STARTUP
+          value: "false"
+        - name: MLSERVER_MODEL_NAME
+          value: dummy-model
+        - name: MLSERVER_HOST
+          value: "127.0.0.1"
+        - name: MLSERVER_GRPC_MAX_MESSAGE_LENGTH
+          value: "-1"
+      resources:
+        requests:
+          cpu: 500m
+          memory: 1Gi
+        limits:
+          cpu: "5"
+          memory: 1Gi
+  builtInAdapter:
+    serverType: mlserver
+    runtimeManagementPort: 8001
+    memBufferBytes: 134217728
+    modelLoadingTimeoutMillis: 90000
+EOF
+```
+
+Verify the available `ServingRuntime`, including the custom one:
+
+```shell
+kubectl get servingruntimes
+
+NAME                  DISABLED   MODELTYPE      CONTAINERS   AGE
+mlserver-1.x                     sklearn        mlserver     10m
+my-custom-model-0.x              custom_model   mlserver     10m
+ovms-1.x                         openvino_ir    ovms         10m
+torchserve-0.x                   pytorch-mar    torchserve   10m
+triton-2.x                       keras          triton       10m
+```
+
+## 4. Deploy the InferenceService using the custom ServingRuntime
+
+```shell
+kubectl apply -f - <<EOF
+---
+apiVersion: serving.kserve.io/v1beta1
+kind: InferenceService
+metadata:
+  name: sklearn-pvc-example
+  annotations:
+    serving.kserve.io/deploymentMode: ModelMesh
+spec:
+  predictor:
+    model:
+      modelFormat:
+        name: custom-model
+      runtime: my-custom-model-0.x
+      storage:
+        parameters:
+          type: pvc
+          name: my-models-pvc
+        path: mnist-svm.joblib
+EOF
+```
+
+After a few seconds, this InferenceService named `sklearn-pvc-example` should be ready:
+
+```shell
+kubectl get isvc
+NAME                  URL                                               READY   PREV   LATEST   PREVROLLEDOUTREVISION   LATESTREADYREVISION   AGE
+sklearn-pvc-example   grpc://modelmesh-serving.modelmesh-serving:8033   True                                                                  69s
+```
+
+## 5. Run an inference request for this InferenceService
+
+Firstly, set up a port-forward to facilitate REST requests:
+
+```shell
+kubectl port-forward --address 0.0.0.0 service/modelmesh-serving 8008 &
+
+# [1] running kubectl port-forward in the background
+# Forwarding from 0.0.0.0:8008 -> 8008
+```
+
+Performing an inference request to the SKLearn MNIST model via `curl`. Make sure the `MODEL_NAME` variable is set correctly.
+
+```shell
+MODEL_NAME="sklearn-pvc-example"
+curl -s -X POST -k "http://localhost:8008/v2/models/${MODEL_NAME}/infer" -d '{"inputs": [{ "name": "predict", "shape": [1, 64], "datatype": "FP32", "data": [0.0, 0.0, 1.0, 11.0, 14.0, 15.0, 3.0, 0.0, 0.0, 1.0, 13.0, 16.0, 12.0, 16.0, 8.0, 0.0, 0.0, 8.0, 16.0, 4.0, 6.0, 16.0, 5.0, 0.0, 0.0, 5.0, 15.0, 11.0, 13.0, 14.0, 0.0, 0.0, 0.0, 0.0, 2.0, 12.0, 16.0, 13.0, 0.0, 0.0, 0.0, 0.0, 0.0, 13.0, 16.0, 16.0, 6.0, 0.0, 0.0, 0.0, 0.0, 16.0, 16.0, 16.0, 7.0, 0.0, 0.0, 0.0, 0.0, 11.0, 13.0, 12.0, 1.0, 0.0]}]}' | jq .
+{
+  "model_name": "sklearn-pvc-example__isvc-72fbffc584",
+  "outputs": [
+    {
+      "name": "predict",
+      "datatype": "INT64",
+      "shape": [1],
+      "data": [8]
+    }
+  ]
+}
+```
+
+> **Note**: `jq` is optional, it is used to format the output of the InferenceService.
+
+To delete the resources created in this example, run the following commands:
+
+```shell
+kubectl delete isvc "sklearn-pvc-example"
+kubectl delete pod "pvc-access"
+kubectl delete pvc "my-models-pvc"
+```

docs/examples/python-custom-runtime/custom-model/Dockerfile

@@ -0,0 +1,35 @@
+FROM python:3.9.13
+# ENV LANG C.UTF-8
+
+COPY requirements.txt ./requirements.txt
+RUN pip3 install --no-cache-dir -r requirements.txt
+
+# The custom `MLModel` implementation should be on the Python search path
+# instead of relying on the working directory of the image. If using a
+# single-file module, this can be accomplished with:
+COPY --chown=${USER} ./custom_model.py /opt/custom_model.py
+ENV PYTHONPATH=/opt/
+WORKDIR /opt
+
+# environment variables to be compatible with ModelMesh Serving
+#  these can also be set in the ServingRuntime, but this is recommended for
+#  consistency when building and testing
+# reference: https://mlserver.readthedocs.io/en/latest/reference/settings.html
+ENV MLSERVER_MODELS_DIR=/models/_mlserver_models \
+    MLSERVER_GRPC_PORT=8001 \
+    MLSERVER_HTTP_PORT=8002 \
+    MLSERVER_METRICS_PORT=8082 \
+    MLSERVER_LOAD_MODELS_AT_STARTUP=false \
+    MLSERVER_DEBUG=false \
+    MLSERVER_PARALLEL_WORKERS=1 \
+    MLSERVER_GRPC_MAX_MESSAGE_LENGTH=33554432 \
+    # https://github.com/SeldonIO/MLServer/pull/748
+    MLSERVER__CUSTOM_GRPC_SERVER_SETTINGS='{"grpc.max_metadata_size": "32768"}' \
+    MLSERVER_MODEL_NAME=dummy-model
+
+# With this setting, the implementation field is not required in the model
+# settings which eases integration by allowing the built-in adapter to generate
+# a basic model settings file
+ENV MLSERVER_MODEL_IMPLEMENTATION=custom_model.CustomMLModel
+
+CMD mlserver start $MLSERVER_MODELS_DIR

docs/examples/python-custom-runtime/custom-model/custom_model.py

@@ -0,0 +1,61 @@
+import os
+from os.path import exists
+from typing import Dict, List
+from mlserver import MLModel
+from mlserver.utils import get_model_uri
+from mlserver.types import InferenceRequest, InferenceResponse, ResponseOutput, Parameters
+from mlserver.codecs import DecodedParameterName
+from joblib import load
+
+import logging
+import numpy as np
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+_to_exclude = {
+    "parameters": {DecodedParameterName, "headers"},
+    'inputs': {"__all__": {"parameters": {DecodedParameterName, "headers"}}}
+}
+
+WELLKNOWN_MODEL_FILENAMES = ["mnist-svm.joblib"]
+
+
+class CustomMLModel(MLModel):  # pylint:disable=c-extension-no-member
+    async def load(self) -> bool:
+        model_uri = await get_model_uri(self._settings, wellknown_filenames=WELLKNOWN_MODEL_FILENAMES)
+        logger.info("Model load URI: {model_uri}")
+        if exists(model_uri):
+            logger.info(f"Loading MNIST model from {model_uri}")
+            self._model = load(model_uri)
+            logger.info("Model loaded successfully")
+        else:
+            logger.info(f"Model not exist in {model_uri}")
+            # raise FileNotFoundError(model_uri)
+            self.ready = False
+            return self.ready
+
+        self.ready = True
+        return self.ready
+
+    async def predict(self, payload: InferenceRequest) -> InferenceResponse:
+        input_data = [input_data.data for input_data in payload.inputs]
+        input_name = [input_data.name for input_data in payload.inputs]
+        input_data_array = np.array(input_data)
+        result = self._model.predict(input_data_array) 
+        predictions = np.array(result)
+
+        logger.info(f"Predict result is: {result}")
+        return InferenceResponse(
+            id=payload.id,
+            model_name = self.name,
+            model_version = self.version,
+            outputs = [
+                ResponseOutput(
+                    name = str(input_name[0]),
+                    shape = predictions.shape,
+                    datatype = "INT64",
+                    data=predictions.tolist(),
+                )
+            ],
+        )

docs/examples/python-custom-runtime/custom-model/requirements.txt

@@ -0,0 +1,3 @@
+mlserver==1.3.2
+scikit-learn==0.24.2
+joblib==1.0.1