Ask or search…


How to deploy models to Kubernetes with KServe
This is an older version of the ZenML documentation. To read and view the latest version please visit this up-to-date URL.
The KServe Model Deployer is one of the available flavors of the Model Deployer stack component. Provided with the MLflow and Seldon Core integration it can be used to deploy and manage models on a inference server running on top of a Kubernetes cluster.

When to use it?

KServe is a Kubernetes-based model inference platform built for highly scalable deployment use cases. It provides a standardized inference protocol across ML frameworks while supporting a serverless architecture with autoscaling including Scale to Zero on GPUs. KServe uses a simple and pluggable production serving architecture for production ML serving that includes prediction, pre-/post-processing, monitoring and explainability.
KServe encapsulates the complexity of autoscaling, networking, health checking, and server configuration to bring cutting edge serving features like GPU Autoscaling, Scale to Zero, and Canary Rollouts to your ML deployments. It enables a simple, pluggable, and complete story for Production ML Serving including prediction, pre-processing, post-processing and explainability. KServe is being used across various organizations.
You should use the KServe Model Deployer:
  • If you are looking to deploy your model with an advanced Model Inference Platform with Kubernetes, built for highly scalable use cases.
  • If you want to handle the lifecycle of the deployed model with no downtime, with possibility of scaling to zero on GPUs.
  • Looking for out-of-the-box model serving runtimes that are easy to use and easy to deploy model from the majority of frameworks.
  • If you want more advanced deployment strategies like A/B testing, canary deployments, ensembles and transformers.
  • if you want to overcome the model deployment Scalability problems. Read more about KServe Multi Model Serving or ModelMesh .
If you are looking for a more easy way to deploy your models locally, you can use the MLflow Model Deployer flavor.

How to deploy it?

ZenML provides a KServe flavor build on top of the KServe Integration to allow you to deploy and use your models in a production-grade environment. In order to use the integration you need to install it on your local machine to be able to register the KServe Model deployer with ZenML and add it to your stack:
zenml integration install kserve -y
To deploy and make use of the KServe integration we need to have the following prerequisites:
  1. 1.
    access to a Kubernetes cluster. The example accepts a --kubernetes-context command line argument. This Kubernetes context needs to point to the Kubernetes cluster where KServe model servers will be deployed. If the context is not explicitly supplied to the example, it defaults to using the locally active context. You can find more information about setup and usage of the Kubernetes cluster in the ZenML Cloud Guide
  2. 2.
    KServe needs to be preinstalled and running in the target Kubernetes cluster. Check out the KServe Serverless installation Guide.
  3. 3.
    models deployed with KServe need to be stored in some form of persistent shared storage that is accessible from the Kubernetes cluster where KServe is installed (e.g. AWS S3, GCS, Azure Blob Storage, etc.). You can use one of the supported remote storage flavors to store your models as part of your stack
Since the KServe Model Deployer is interacting with the KServe model serving Platform deployed on a Kubernetes cluster, you need to provide a set of configuration parameters. These parameters are:
  • kubernetes_context: the Kubernetes context to use to contact the remote KServe installation. If not specified, the current configuration is used. Depending on where the KServe model deployer is being used
  • kubernetes_namespace: the Kubernetes namespace where the KServe deployment servers are provisioned and managed by ZenML. If not specified, the namespace set in the current configuration is used.
  • base_url: the base URL of the Kubernetes ingress used to expose the KServe deployment servers.
  • secret: the name of a ZenML secret containing the credentials used by KServe storage initializers to authenticate to the Artifact Store
Configuring KServe in a Kubernetes cluster can be a complex and error prone process, We provide a simple start guide on how to configure and setup KServe on your Kubernetes cluster, you can find it here we have also provided a set of of Terraform-based recipes to quickly provision popular combinations of MLOps tools. More information about these recipes can be found in the Open Source MLOps Stack Recipes

Managing KServe Credentials

The KServe model servers need to access the Artifact Store in the ZenML stack to retrieve the model artifacts. This usually involve passing some credentials to the KServe model servers required to authenticate with the Artifact Store. In ZenML, this is done by creating a ZenML secret with the proper credentials and configuring the KServe Model Deployer stack component to use it, by passing the --secret argument to the CLI command used to register the model deployer. We've already done the latter, now all that is left to do is to configure the s3-store ZenML secret specified before as a KServe Model Deployer configuration attribute with the credentials needed by KServe to access the artifact store.
There are built-in secret schemas that the KServe integration provides which can be used to configure credentials for the 3 main types of Artifact Stores supported by ZenML: S3, GCS and Azure.
you can use kserve_s3 for AWS S3 or kserve_gs for GCS and kserve_az for Azure. To read more about secrets, secret schemas and how they are used in ZenML, please refer to the Secrets Manager.
The recommended way to pass the credentials to the KServe model deployer is to use a file that contains the credentials. You can achieve this by adding the @ followed by the path to the file to the --credentials argument. (e.g. --credentials @/path/to/credentials.json)
The following is an example of registering an GS secret with the KServe model deployer:
$ zenml secrets-manager secret register -s kserve_gs kserve_secret \
--namespace="zenml-workloads" \
--credentials="@~/sa-deployment-temp.json" \
┃ storage_type │ *** ┃
┃ namespace │ *** ┃
┃ credentials │ *** ┃
$ zenml secrets-manager secret get kserve_secret
┃ storage_type │ GCS ┃
┃ namespace │ kserve-test ┃
┃ credentials │ ~/sa-deployment-temp.json ┃
For more information and a full list of configurable attributes of the KServe secret schemas, check out the API Docs.

How do you use it?

We can register the model deployer and use it in our active stack:
zenml model-deployer register kserve_gke --flavor=kserve \
--kubernetes_context=gke_zenml-core_us-east1-b_zenml-test-cluster \
--kubernetes_namespace=zenml-workloads \
--base_url=$INGRESS_URL \
# Now we can use the model deployer in our stack
zenml stack update kserve_stack --model-deployer=kserve_gke
As the packaging and preparation of the model artifacts to the right format can be a bit of a challenge, ZenML's KServe Integration comes with a built-in model deployment step that can be used to deploy your models with the minimum of effort.
This step will:
  • Verify if the model is already deployed in the KServe cluster. If not, it will deploy the model.
  • Prepare the model artifacts to the right format for the TF, MLServer runtimes servers.
  • Package, verify and prepare the model artifact for the PyTorch runtime server since it requires additional files.
  • Upload the model artifacts to the Artifact Store.
An example of how to use the model deployment step is shown below.
from zenml.integrations.kserve.services import KServeDeploymentConfig
from zenml.integrations.kserve.steps import (
MODEL_NAME = "mnist-pytorch"
pytorch_model_deployer = kserve_model_deployer_step(
resources={"requests": {"cpu": "200m", "memory": "500m"}},
A concrete example of using the KServe Model Deployer can be found here.
For more information and a full list of configurable attributes of the KServe Model Deployer, check out the API Docs.
The model deployment step are expiremntal good for standard use cases. However, if you need to customize the deployment step, you can always create your own model deployment step. Find more information about model deployment steps in the Model Deployment Steps section.

Custom Model Deployment

While KServe is a good fit for most use cases with the built-in model servers, it is not always the best fit for your custom model deployment use case. For that reason KServe allows you to create your own model server using the KServe ModelServer API where you can customize the predict, the pre- and post-processing functions. With ZenML's KServe Integration, you can create your own custom model deployment code by creating a custom predict function that will be passed to a custom deployment step responsible for preparing a Docker image for the model server.
This custom_predict function should be getting the model and the input data as arguments and returns the output data. ZenML will take care of loading the model into memory, starting the KServe ModelServer that will be responsible for serving the model, and running the predict function.
def pre_process(tensor: torch.Tensor) -> dict:
"""Pre process the data to be used for prediction."""
def post_process(prediction: torch.Tensor) -> str:
"""Pre process the data"""
def custom_predict(
model: Any,
request: dict,
) -> dict:
"""Custom Prediction function.
The custom predict function is the core of the custom deployment. The function
is called by the custom deployment class defined for the serving tool.
The current implementation requires the function to get the model loaded in the memory and
a request with the data to predict.
model (Any): The model to use for prediction.
request: The prediction response of the model is an array-like object.
The prediction in an array-like format (e.g. np.ndarray, List[Any], str, bytes, Dict[str, Any])
Then this custom predict function path can be passed to the custom deployment parameters.
kserve_pytorch_custom_deployment = kserve_custom_model_deployer_step(
resources={"requests": {"cpu": "200m", "memory": "500m"}},
The full code example can be found here.

Advanced Custom Code Deployment with KServe Integration

Before creating your custom model class, you should take a look at the 'Deploy Custom Python Model Server with InferenceService' section of the KServe documentation.
The built-in KServe custom deployment step is a good starting point for deploying your custom models. However if you want to deploy more than the trained model, you can create your own Custom Model Class and a custom step to achieve this.
Example of the custom model class
The built-in KServe custom deployment step responsible for packaging, preparing and deploying to KServe can be found here