Skip to main content
Version: Next

Quick Start

After installing KAITO, you can quickly deploy a phi-4-mini-instruct inference service to get started.

Prerequisites​

  • A Kubernetes cluster with KAITO installed (see Installation)
  • kubectl configured to access your cluster
  • GPU nodes available in your cluster - You have two mutually exclusive options:
    • Bring your own GPU (BYO) nodes: If you already have GPU nodes in your cluster, you can use the preferred nodes approach (shown below). Requires Node Auto Provisioning to be disabled.
    • Auto-provisioning: Set up automatic GPU node provisioning for your cloud provider (see these steps). Cannot be used with BYO nodes.

Deploy Your First Model​

Your model can be deployed either on your own GPU nodes, or KAITO can auto-provision nodes for it to run on. These options are mutually exclusive - you must choose one approach.

Important

BYO nodes and Auto-provisioning are mutually exclusive. If you're using your own GPU nodes, ensure KAITO was installed with --set featureGates.disableNodeAutoProvisioning=true. If you're using auto-provisioning, do not use this flag.

Option 1: Bring your own GPU nodes​

Prerequisites for BYO Nodes

Before using this option, ensure that:

  1. KAITO was installed with Node Auto Provisioning disabled: --set featureGates.disableNodeAutoProvisioning=true
  2. You have existing GPU nodes in your cluster
  3. Your GPU nodes are properly labeled for workload selection

Let's start by deploying a phi-4-mini-instruct model on your existing GPU nodes.

First get the nodes.

kubectl get nodes -l accelerator=nvidia

The output should look similar to this, showing all your GPU nodes.

NAME                                  STATUS   ROLES    AGE     VERSION
gpunp-26695285-vmss000000             Ready    <none>   2d21h   v1.31.9
gpunp-26695285-vmss000001             Ready    <none>   2d21h   v1.31.9

The GPU nodes will need a label in order for a KAITO Workspace to select it. We'll use the label apps=llm-inference for this example. Label the nodes you want to use.

kubectl label node gpunp-26695285-vmss000001 apps=llm-inference
kubectl label node gpunp-26695285-vmss000000 apps=llm-inference

Create a YAML file named phi-4-workspace.yaml with the following content. The label selector here must match the labels you set on the nodes.

phi-4-workspace.yaml
apiVersion: kaito.sh/v1beta1
kind: Workspace
metadata:
  name: workspace-phi-4-mini
resource:
  labelSelector:
    matchLabels:
      apps: llm-inference
inference:
  preset:
    name: phi-4-mini-instruct
note

Make sure that resource.instanceType is empty when using BYO as it's only used for the auto-provisioning scenario.

Apply your configuration to your cluster:

kubectl apply -f phi-4-workspace.yaml

The nodes will be populated with their specs using NVIDIA GPU Feature Discovery, and KAITO will verify that your nodes have the proper memory, # GPUS, etc. to run the model.

Option 2: Auto-provision GPU nodes​

The following cloud providers support auto-provisioning GPU nodes in addition to BYO nodes.

info

If you have not already, follow the steps here to install the gpu-provisioner Helm chart.

Ensure that the chart is ready and pods are running with these steps otherwise the GPU nodes will not be provisioned.

Create a YAML file named phi-4-workspace.yaml with the following content. The instanceType field will specify what nodes will be auto-provisioned instead of only matching existing nodes in the BYO case.

phi-4-workspace.yaml
apiVersion: kaito.sh/v1beta1
kind: Workspace
metadata:
  name: workspace-phi-4-mini
resource:
  instanceType: "Standard_NC6s_v3" # Specifies the node type that will be auto-provisioned.
  labelSelector:
    matchLabels:
      apps: phi-4-mini
inference:
  preset:
    name: phi-4-mini-instruct

Apply your configuration to your cluster:

kubectl apply -f phi-4-workspace.yaml

Monitor Deployment​

Track the workspace status to see when the model has been deployed successfully:

kubectl get workspace workspace-phi-4-mini

When the WORKSPACESUCCEEDED column becomes True, the model has been deployed successfully:

NAME                   INSTANCE                   RESOURCEREADY   INFERENCEREADY   JOBSTARTED   WORKSPACESUCCEEDED   AGE
workspace-phi-4-mini   Standard_NC24ads_A100_v4   True            True                          True                 4h15m
note

The INSTANCE column will default to Standard_NC24ads_A100_v4 if you have not set up auto-provisioning. If you have auto-provisioning configured, it will show the specific instance type used.

Test the Model​

Find the inference service's cluster IP and test it using a temporary curl pod:

# Get the service endpoint
kubectl get svc workspace-phi-4-mini
export CLUSTERIP=$(kubectl get svc workspace-phi-4-mini -o jsonpath="{.spec.clusterIPs[0]}")

# List available models
kubectl run -it --rm --restart=Never curl --image=curlimages/curl -- curl -s http://$CLUSTERIP/v1/models | jq

You should see output similar to:

{
  "object": "list",
  "data": [
    {
      "id": "phi-4-mini-instruct",
      "object": "model",
      "created": 1733370094,
      "owned_by": "vllm",
      "root": "/workspace/vllm/weights",
      "parent": null,
      "max_model_len": 16384
    }
  ]
}

Make an Inference Call​

Now make an inference call using the model:

kubectl run -it --rm --restart=Never curl --image=curlimages/curl -- curl -X POST http://$CLUSTERIP/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "phi-4-mini-instruct",
    "messages": [{"role": "user", "content": "What is kubernetes?"}],
    "max_tokens": 50,
    "temperature": 0
  }'

Next Steps​

🎉 Congratulations! You've successfully deployed and tested your first model with KAITO.

What's Next:

Additional Resources: