feat(operator): add DisaggregatedSet pathway

deploy/helm/charts/platform/components/operator/files/manager-role.yaml

@@ -128,6 +128,18 @@ rules:
   - patch
   - update
   - watch
+- apiGroups:
+  - disaggregatedset.x-k8s.io
+  resources:
+  - disaggregatedsets
+  verbs:
+  - create
+  - delete
+  - get
+  - list
+  - patch
+  - update
+  - watch
 - apiGroups:
   - discovery.k8s.io
   resources:

deploy/operator/config/rbac/role.yaml

@@ -128,6 +128,18 @@ rules:
   - patch
   - update
   - watch
+- apiGroups:
+  - disaggregatedset.x-k8s.io
+  resources:
+  - disaggregatedsets
+  verbs:
+  - create
+  - delete
+  - get
+  - list
+  - patch
+  - update
+  - watch
 - apiGroups:
   - discovery.k8s.io
   resources:

deploy/operator/internal/consts/consts.go

@@ -40,7 +40,8 @@ const (
 
 	KubeLabelDynamoSelector = "nvidia.com/selector"
 
-	KubeAnnotationEnableGrove = "nvidia.com/enable-grove"
+	KubeAnnotationEnableGrove            = "nvidia.com/enable-grove"
+	KubeAnnotationEnableDisaggregatedSet = "nvidia.com/enable-disaggregatedset"
 
 	// KubeAnnotationGroveUpdateStrategy temporarily exposes the Grove
 	// PodCliqueSet update strategy while the long-term DGD API is settled.

deploy/operator/internal/controller/dynamocomponentdeployment_controller.go

@@ -551,61 +551,74 @@ func (r *DynamoComponentDeploymentReconciler) generateLeaderWorkerSet(ctx contex
 	logs := log.FromContext(ctx)
 	logs.Info("Generating LeaderWorkerSet")
 
+	leaderPodTemplateSpec, workerPodTemplateSpec, err := r.renderMultinodePodTemplateSpecs(ctx, opt)
+	if err != nil {
+		return nil, false, err
+	}
+
 	kubeName := leaderWorkerSetName(opt.dynamoComponentDeployment)
 	kubeNs := opt.dynamoComponentDeployment.Namespace
 	labels := dynamo.GetDCDKubeLabels(opt.dynamoComponentDeployment)
-
 	if labels == nil {
 		labels = make(map[string]string)
 	}
-	podLabels, err := r.getDCDWorkloadPodLabels(ctx, opt.dynamoComponentDeployment)
-	if err != nil {
-		return nil, false, err
+
+	desiredReplicas := int32(1)
+	if opt.dynamoComponentDeployment.Spec.Replicas != nil {
+		desiredReplicas = *opt.dynamoComponentDeployment.Spec.Replicas
 	}
+	groupSize := opt.dynamoComponentDeployment.GetNumberOfNodes()
 
-	leaderWorkerSet := &leaderworkersetv1.LeaderWorkerSet{
+	return &leaderworkersetv1.LeaderWorkerSet{
 		ObjectMeta: metav1.ObjectMeta{
 			Name:      kubeName,
 			Namespace: kubeNs,
 			Labels:    labels,
 		},
-	}
+		Spec: leaderworkersetv1.LeaderWorkerSetSpec{
+			Replicas:      &desiredReplicas,
+			StartupPolicy: leaderworkersetv1.LeaderCreatedStartupPolicy,
+			LeaderWorkerTemplate: leaderworkersetv1.LeaderWorkerTemplate{
+				LeaderTemplate: leaderPodTemplateSpec,
+				WorkerTemplate: *workerPodTemplateSpec,
+				Size:           &groupSize,
+			},
+		},
+	}, false, nil
+}
 
-	leaderPodLabels := make(map[string]string)
-	for k, v := range podLabels {
-		leaderPodLabels[k] = v
-	}
-	leaderPodTemplateSpec, err := r.generateLeaderPodTemplateSpec(ctx, opt, leaderPodLabels)
+// renderMultinodePodTemplateSpecs is the shared multinode render path used by
+// both the LWS pathway and the DisaggregatedSet pathway. Returning the two
+// pod templates (rather than a fully composed LWS) keeps downstream callers
+// (DS, future templating work) free to compose their own role structure.
+func (r *DynamoComponentDeploymentReconciler) renderMultinodePodTemplateSpecs(
+	ctx context.Context,
+	opt generateResourceOption,
+) (*corev1.PodTemplateSpec, *corev1.PodTemplateSpec, error) {
+	podLabels, err := r.getDCDWorkloadPodLabels(ctx, opt.dynamoComponentDeployment)
 	if err != nil {
-		return nil, false, errors.Wrap(err, "generateLeaderWorkerSet: failed to generate leader pod template")
+		return nil, nil, err
 	}
 
-	workerPodLabels := make(map[string]string)
+	leaderLabels := make(map[string]string, len(podLabels))
 	for k, v := range podLabels {
-		workerPodLabels[k] = v
+		leaderLabels[k] = v
 	}
-	workerPodTemplateSpec, err := r.generateWorkerPodTemplateSpec(ctx, opt, workerPodLabels)
+	leaderPodTemplateSpec, err := r.generateLeaderPodTemplateSpec(ctx, opt, leaderLabels)
 	if err != nil {
-		return nil, false, errors.Wrap(err, "generateLeaderWorkerSet: failed to generate worker pod template")
+		return nil, nil, err
 	}
 
-	desiredReplicas := int32(1)
-	if opt.dynamoComponentDeployment.Spec.Replicas != nil {
-		desiredReplicas = *opt.dynamoComponentDeployment.Spec.Replicas
+	workerLabels := make(map[string]string, len(podLabels))
+	for k, v := range podLabels {
+		workerLabels[k] = v
 	}
-	groupSize := opt.dynamoComponentDeployment.GetNumberOfNodes()
-
-	leaderWorkerSet.Spec = leaderworkersetv1.LeaderWorkerSetSpec{
-		Replicas:      &desiredReplicas,
-		StartupPolicy: leaderworkersetv1.LeaderCreatedStartupPolicy,
-		LeaderWorkerTemplate: leaderworkersetv1.LeaderWorkerTemplate{
-			LeaderTemplate: leaderPodTemplateSpec,
-			WorkerTemplate: *workerPodTemplateSpec,
-			Size:           &groupSize,
-		},
+	workerPodTemplateSpec, err := r.generateWorkerPodTemplateSpec(ctx, opt, workerLabels)
+	if err != nil {
+		return nil, nil, err
 	}
 
-	return leaderWorkerSet, false, nil
+	return leaderPodTemplateSpec, workerPodTemplateSpec, nil
 }
 
 // getDCDWorkloadPodLabels keeps LWS pod labels aligned with the workload