Installation

This guide describes how to install the Training Operator on your Kubernetes cluster. The Training Operator is a lightweight Kubernetes controller that orchestrates the appropriate Kubernetes workloads to perform distributed ML training and fine-tuning.

Prerequisites

These are the minimal requirements to install the Training Operator:

• Kubernetes >= 1.27
• kubectl >= 1.27
• Python >= 3.7

Installing the Training Operator

You need to install the Training Operator control plane and Python SDK to create training jobs.

Installing the Control Plane

You can skip these steps if you have already installed Kubeflow platform using manifests or package distributions. The Kubeflow platform includes the Training Operator.

You can install the Training Operator as a standalone component.

Run the following command to install the stable release of the Training Operator control plane: v1.8.1

kubectl apply --server-side -k "github.com/kubeflow/training-operator.git/manifests/overlays/standalone?ref=v1.8.1"

Run the following command to install the latest changes of Training Operator control plane:

kubectl apply --server-side -k "github.com/kubeflow/training-operator.git/manifests/overlays/standalone?ref=master"

After installing it, you can verify that Training Operator controller is running as follows:

$ kubectl get pods -n kubeflow

NAME                                             READY   STATUS    RESTARTS   AGE
training-operator-658c68d697-46zmn               1/1     Running   0          90s

Run this command to check installed Kubernetes CRDs for each supported ML framework:

$ kubectl get crd

mpijobs.kubeflow.org                                     2023-06-09T00:31:07Z
mxjobs.kubeflow.org                                      2023-06-09T00:31:05Z
paddlejobs.kubeflow.org                                  2023-06-09T00:31:09Z
pytorchjobs.kubeflow.org                                 2023-06-09T00:31:06Z
tfjobs.kubeflow.org                                      2023-06-09T00:31:04Z
xgboostjobs.kubeflow.org                                 2023-06-09T00:31:04Z

Installing the Python SDK

The Training Operator implements a Python SDK to simplify creation of distributed training and fine-tuning jobs for Data Scientists.

Run the following command to install the latest stable release of the Training SDK:

pip install -U kubeflow-training

Run the following command to install the latest changes of Training SDK:

pip install git+https://github.com/kubeflow/training-operator.git@master#subdirectory=sdk/python

Otherwise, you can also install the Training SDK using the specific GitHub commit, for example:

pip install git+https://github.com/kubeflow/training-operator.git@7345e33b333ba5084127efe027774dd7bed8f6e6#subdirectory=sdk/python

Install the Python SDK with Fine-Tuning Capabilities

If you want to use the train API for LLM fine-tuning with the Training Operator, install the Python SDK with the additional packages from HuggingFace:

pip install -U "kubeflow-training[huggingface]"

Next steps

Run your first Training Operator Job by following the Getting Started guide.