#Integrate with LeaderWorkerSet

This page shows how to use Alauda Build of Kueue to manage LeaderWorkerSet (LWS) workloads. A LeaderWorkerSet is a Kubernetes API that enables a common deployment pattern where a single leader coordinates multiple workers.

LeaderWorkerSet is particularly useful for distributed inference and training scenarios where a leader process manages the execution of multiple worker processes.

#Prerequisites

• You have installed the Alauda Build of Kueue.
• You have installed the LeaderWorkerSet controller.
• The LeaderWorkerSet framework is enabled in the Kueue configuration.
• The Alauda Container Platform Web CLI has communication with your cluster.
• You have created a ClusterQueue, ResourceFlavor, and LocalQueue.

#Procedure

1. Create a LeaderWorkerSet resource with the kueue.x-k8s.io/queue-name label:

   apiVersion: leaderworkerset.x-k8s.io/v1
   kind: LeaderWorkerSet
   metadata:
     name: lws-training
     namespace: team-ml
     labels:
       kueue.x-k8s.io/queue-name: team-ml-queue
   spec:
     replicas: 2
     leaderWorkerTemplate:
       size: 3
       restartPolicy: RecreateGroupOnPodRestart
       leaderTemplate:
         metadata: {}
         spec:
           containers:
           - name: leader
             image: pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime
             resources:
               requests:
                 cpu: "2"
                 memory: "4Gi"
             command: ["python", "leader.py"]
       workerTemplate:
         metadata: {}
         spec:
           containers:
           - name: worker
             image: pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime
             resources:
               requests:
                 cpu: "2"
                 memory: "4Gi"
                 nvidia.com/gpu: "1"
             command: ["python", "worker.py"]
     rolloutStrategy:
       type: RollingUpdate
       rollingUpdateConfiguration:
         maxSurge: 1
         maxUnavailable: 1
     startupPolicy: LeaderCreated
   1. kueue.x-k8s.io/queue-name: Specifies the LocalQueue that manages this LeaderWorkerSet. Kueue admits all groups together.
   2. replicas: 2: Creates 2 leader-worker groups, each managed independently.
   3. size: 3: Each group consists of 1 leader and 2 workers (size = total pods including leader).

2. Apply the LeaderWorkerSet:

   kubectl apply -f lws-training.yaml

3. Monitor the admission:

   kubectl get workloads -n team-ml

4. Check the LeaderWorkerSet status:

   kubectl get leaderworkersets lws-training -n team-ml

5. View individual pods:

   kubectl get pods -n team-ml -l leaderworkerset.sigs.k8s.io/name=lws-training