feat(operator): add experimental grove.forceScalingGroup for single-node components

deploy/operator/api/v1alpha1/conversion_field_coverage_test.go

@@ -58,6 +58,7 @@ DynamoComponentDeploymentSpec.experimental.gpuMemoryService.extraClientContainer
 DynamoComponentDeploymentSpec.experimental.gpuMemoryService.extraClientPods.name
 DynamoComponentDeploymentSpec.experimental.gpuMemoryService.extraClientPods.podTemplate
 DynamoComponentDeploymentSpec.experimental.gpuMemoryService.mode
+DynamoComponentDeploymentSpec.experimental.grove.forceScalingGroup
 DynamoComponentDeploymentSpec.frontendSidecar
 DynamoComponentDeploymentSpec.globalDynamoNamespace
 DynamoComponentDeploymentSpec.minAvailable
@@ -117,6 +118,7 @@ DynamoGraphDeploymentSpec.components.experimental.gpuMemoryService.extraClientCo
 DynamoGraphDeploymentSpec.components.experimental.gpuMemoryService.extraClientPods.name
 DynamoGraphDeploymentSpec.components.experimental.gpuMemoryService.extraClientPods.podTemplate
 DynamoGraphDeploymentSpec.components.experimental.gpuMemoryService.mode
+DynamoGraphDeploymentSpec.components.experimental.grove.forceScalingGroup
 DynamoGraphDeploymentSpec.components.frontendSidecar
 DynamoGraphDeploymentSpec.components.globalDynamoNamespace
 DynamoGraphDeploymentSpec.components.minAvailable

deploy/operator/api/v1alpha1/dynamocomponentdeployment_conversion_test.go

@@ -623,6 +623,48 @@ func TestDCD_RoundTrip_Experimental(t *testing.T) {
 	}
 }
 
+// The grove block has no v1alpha1 representation and must survive the spoke
+// round-trip both alone (whole hub-only block preserved) and alongside
+// alpha-representable fields (sparse preservation merged back).
+func TestDCD_RoundTrip_ExperimentalGrove(t *testing.T) {
+	tests := []struct {
+		name         string
+		experimental *v1beta1.ExperimentalSpec
+	}{
+		{
+			name: "grove.forceScalingGroup only",
+			experimental: &v1beta1.ExperimentalSpec{
+				Grove: &v1beta1.GroveSpec{ForceScalingGroup: true},
+			},
+		},
+		{
+			name: "grove.forceScalingGroup alongside alpha-representable GMS",
+			experimental: &v1beta1.ExperimentalSpec{
+				GPUMemoryService: &v1beta1.GPUMemoryServiceSpec{Mode: v1beta1.GMSModeIntraPod},
+				Grove:            &v1beta1.GroveSpec{ForceScalingGroup: true},
+			},
+		},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := &v1beta1.DynamoComponentDeployment{
+				ObjectMeta: metav1.ObjectMeta{Name: "exp-grouping", Namespace: "ns"},
+				Spec: v1beta1.DynamoComponentDeploymentSpec{
+					DynamoComponentDeploymentSharedSpec: v1beta1.DynamoComponentDeploymentSharedSpec{
+						ComponentName: "exp-grouping",
+						ComponentType: v1beta1.ComponentTypeWorker,
+						Experimental:  tt.experimental.DeepCopy(),
+					},
+				},
+			}
+			got := dcdRoundTripFromV1beta1(t, src)
+			if diff := cmp.Diff(src, got, cmpopts.EquateEmpty()); diff != "" {
+				t.Errorf("round-trip mismatch (-want +got):\n%s", diff)
+			}
+		})
+	}
+}
+
 func TestDCD_ExperimentalModeValuesAreValidForIntermediateVersion(t *testing.T) {
 	alpha := &DynamoComponentDeployment{
 		ObjectMeta: metav1.ObjectMeta{Name: "alpha-enums", Namespace: "ns"},

deploy/operator/api/v1alpha1/shared_spec_conversion.go

@@ -596,6 +596,10 @@ func saveSharedHubOnlySpec(src *v1beta1.DynamoComponentDeploymentSharedSpec, con
 	}
 	if experimentalIsHubOnlyShape(src.Experimental) {
 		save.Experimental = src.Experimental.DeepCopy()
+	} else if src.Experimental != nil && src.Experimental.Grove != nil {
+		// The grove block has no v1alpha1 representation; preserve it sparsely
+		// when the rest of the experimental block converts to alpha fields.
+		save.Experimental = &v1beta1.ExperimentalSpec{Grove: src.Experimental.Grove.DeepCopy()}
 	}
 	return nil
 }
@@ -1618,6 +1622,11 @@ func restoreSharedHubOnlyFields(dst, preserved *v1beta1.DynamoComponentDeploymen
 	restoreSharedHubOnlyFrontendSidecar(dst, preserved)
 	if dst.Experimental == nil && experimentalIsHubOnlyShape(preserved.Experimental) {
 		dst.Experimental = preserved.Experimental.DeepCopy()
+	} else if dst.Experimental != nil && preserved.Experimental != nil &&
+		dst.Experimental.Grove == nil && preserved.Experimental.Grove != nil {
+		// The experimental block was rebuilt from alpha fields (GMS, failover,
+		// checkpoint); merge back the sparsely preserved hub-only grove block.
+		dst.Experimental.Grove = preserved.Experimental.Grove.DeepCopy()
 	}
 	return nil
 }

deploy/operator/api/v1beta1/common.go

@@ -164,6 +164,22 @@ const (
 	GMSModeInterPod GPUMemoryServiceMode = "InterPod"
 )
 
+// GroveSpec groups experimental Grove-specific rendering options.
+type GroveSpec struct {
+	// forceScalingGroup opts a single-node component into rendering as a
+	// PodCliqueScalingGroup with one single-pod PodClique per replica.
+	// Scaling changes the scaling-group replica count. The first
+	// `minAvailable` replicas join the deployment's base PodGang together
+	// with its other base workloads; each replica beyond `minAvailable`
+	// gets its own PodGang, gang-scheduled separately from the rest of the
+	// deployment. `false` or omitted means automatic selection (multi-node
+	// and inter-pod GMS components use a scaling group, other single-node
+	// components a standalone PodClique), not "force PodClique".
+	// Immutable after creation.
+	// +optional
+	ForceScalingGroup bool `json:"forceScalingGroup,omitempty"`
+}
+
 // ExperimentalSpec groups opt-in preview features whose API shape and behavior
 // may change in breaking ways between v1beta1 releases (including disappearing
 // without a name-preserving graduation path). Fields placed under
@@ -183,6 +199,10 @@ type ExperimentalSpec struct {
 	// +optional
 	Failover *FailoverSpec `json:"failover,omitempty"`
 
+	// grove groups Grove-specific rendering options.
+	// +optional
+	Grove *GroveSpec `json:"grove,omitempty"`
+
 	// checkpoint configures container-image snapshotting and restore for
 	// this component. Set `checkpoint.enabled: true` to opt in. Without
 	// checkpointRef, the DGD controller creates a DGD-scoped DynamoCheckpoint

deploy/operator/api/v1beta1/dynamocomponentdeployment_types.go

@@ -133,8 +133,10 @@ type DynamoComponentDeploymentSharedSpec struct {
 	// +optional
 	Replicas *int32 `json:"replicas,omitempty"`
 
-	// minAvailable maps to Grove PodClique minAvailable for single-node and
-	// Grove PodCliqueScalingGroup minAvailable for multi-node components.
+	// minAvailable maps to Grove PodCliqueScalingGroup minAvailable for
+	// components rendered as a scaling group (multi-node, inter-pod GMS, or
+	// `experimental.grove.forceScalingGroup`; see `UsesPCSG`) and to Grove
+	// PodClique minAvailable for all other single-node components.
 	// This field determines 1) the minimum number of replicas guaranteed to be
 	// gang-scheduled, and 2) when violating minAvailable replicas triggers gang
 	// termination.
@@ -327,6 +329,19 @@ func (s *DynamoComponentDeploymentSharedSpec) IsInterPodGMSEnabled() bool {
 		s.Experimental.GPUMemoryService.Mode == GMSModeInterPod
 }
 
+// IsGroveScalingGroupForced reports whether the ScalingGroup layout is explicitly requested.
+func (s *DynamoComponentDeploymentSharedSpec) IsGroveScalingGroupForced() bool {
+	return s.Experimental != nil &&
+		s.Experimental.Grove != nil &&
+		s.Experimental.Grove.ForceScalingGroup
+}
+
+// UsesPCSG reports whether Grove renders this component as a
+// PodCliqueScalingGroup rather than a standalone PodClique.
+func (s *DynamoComponentDeploymentSharedSpec) UsesPCSG() bool {
+	return s.GetNumberOfNodes() > 1 || s.IsInterPodGMSEnabled() || s.IsGroveScalingGroupForced()
+}
+
 // IsInterPodFailoverEnabled reports whether inter-pod GMS failover is configured.
 func (s *DynamoComponentDeploymentSharedSpec) IsInterPodFailoverEnabled() bool {
 	return s.Experimental != nil &&

deploy/operator/config/crd/bases/nvidia.com_dynamocomponentdeployments.yaml

@@ -12186,6 +12186,23 @@ spec:
                           rule: '!has(self.extraClientPods) || size(self.extraClientPods) == 0 || self.mode == ''InterPod'''
                         - message: extraClientPods is reserved for inter-pod GMS and is not implemented yet
                           rule: '!has(self.extraClientPods) || size(self.extraClientPods) == 0'
+                    grove:
+                      description: grove groups Grove-specific rendering options.
+                      properties:
+                        forceScalingGroup:
+                          description: |-
+                            forceScalingGroup opts a single-node component into rendering as a
+                            PodCliqueScalingGroup with one single-pod PodClique per replica.
+                            Scaling changes the scaling-group replica count. The first
+                            `minAvailable` replicas join the deployment's base PodGang together
+                            with its other base workloads; each replica beyond `minAvailable`
+                            gets its own PodGang, gang-scheduled separately from the rest of the
+                            deployment. `false` or omitted means automatic selection (multi-node
+                            and inter-pod GMS components use a scaling group, other single-node
+                            components a standalone PodClique), not "force PodClique".
+                            Immutable after creation.
+                          type: boolean
+                      type: object
                   type: object
                 frontendSidecar:
                   description: |-
@@ -12208,8 +12225,10 @@ spec:
                   type: boolean
                 minAvailable:
                   description: |-
-                    minAvailable maps to Grove PodClique minAvailable for single-node and
-                    Grove PodCliqueScalingGroup minAvailable for multi-node components.
+                    minAvailable maps to Grove PodCliqueScalingGroup minAvailable for
+                    components rendered as a scaling group (multi-node, inter-pod GMS, or
+                    `experimental.grove.forceScalingGroup`; see `UsesPCSG`) and to Grove
+                    PodClique minAvailable for all other single-node components.
                     This field determines 1) the minimum number of replicas guaranteed to be
                     gang-scheduled, and 2) when violating minAvailable replicas triggers gang
                     termination.

deploy/operator/config/crd/bases/nvidia.com_dynamographdeployments.yaml

@@ -12644,6 +12644,23 @@ spec:
                                 rule: '!has(self.extraClientPods) || size(self.extraClientPods) == 0 || self.mode == ''InterPod'''
                               - message: extraClientPods is reserved for inter-pod GMS and is not implemented yet
                                 rule: '!has(self.extraClientPods) || size(self.extraClientPods) == 0'
+                          grove:
+                            description: grove groups Grove-specific rendering options.
+                            properties:
+                              forceScalingGroup:
+                                description: |-
+                                  forceScalingGroup opts a single-node component into rendering as a
+                                  PodCliqueScalingGroup with one single-pod PodClique per replica.
+                                  Scaling changes the scaling-group replica count. The first
+                                  `minAvailable` replicas join the deployment's base PodGang together
+                                  with its other base workloads; each replica beyond `minAvailable`
+                                  gets its own PodGang, gang-scheduled separately from the rest of the
+                                  deployment. `false` or omitted means automatic selection (multi-node
+                                  and inter-pod GMS components use a scaling group, other single-node
+                                  components a standalone PodClique), not "force PodClique".
+                                  Immutable after creation.
+                                type: boolean
+                            type: object
                         type: object
                       frontendSidecar:
                         description: |-
@@ -12666,8 +12683,10 @@ spec:
                         type: boolean
                       minAvailable:
                         description: |-
-                          minAvailable maps to Grove PodClique minAvailable for single-node and
-                          Grove PodCliqueScalingGroup minAvailable for multi-node components.
+                          minAvailable maps to Grove PodCliqueScalingGroup minAvailable for
+                          components rendered as a scaling group (multi-node, inter-pod GMS, or
+                          `experimental.grove.forceScalingGroup`; see `UsesPCSG`) and to Grove
+                          PodClique minAvailable for all other single-node components.
                           This field determines 1) the minimum number of replicas guaranteed to be
                           gang-scheduled, and 2) when violating minAvailable replicas triggers gang
                           termination.

deploy/operator/internal/controller/dgd_grove_restart_progress_resolver.go

@@ -82,7 +82,7 @@ func (r *groveRestartProgressResolver) Resolve(
 		// Any component represented by a PodCliqueScalingGroup must use the
 		// PCSG readiness path. Read failures conservatively keep the component
 		// in progress; authoritative readiness returns the error separately.
-		if component.GetNumberOfNodes() > 1 || component.IsInterPodGMSEnabled() {
+		if component.UsesPCSG() {
 			isReady, reason, _, _, _ = dynamo.CheckPCSGReady(
 				ctx,
 				r.reader,

deploy/operator/internal/controller/dgd_grove_scaler.go

@@ -70,7 +70,7 @@ func (s *groveScaler) Reconcile(
 			replicas = 0
 		}
 
-		usesPCSG := component.GetNumberOfNodes() > 1 || component.IsInterPodGMSEnabled()
+		usesPCSG := component.UsesPCSG()
 		resourceName := dynamo.GroveComponentResourceName(dgd, componentName)
 		resourceKind := "PodClique"
 		gvr := consts.PodCliqueGVR

deploy/operator/internal/dynamo/graph.go

@@ -1161,6 +1161,8 @@ const (
 // multiple ServiceRoles depending on the deployment topology:
 //
 //   - single-node, no GMS: 1 role (RoleMain)
+//   - single-node, forceScalingGroup: 1 role (RoleMain with a single pod;
+//     the PCSG replica count carries the horizontal scale)
 //   - multinode, no GMS:    2 roles (RoleLeader + RoleWorker)
 //   - single-node, inter-pod GMS: 1 engine PCLQ (replicated) + 1 RoleGMS
 //     weight-server PCLQ
@@ -1218,6 +1220,8 @@ func expandRolesForComponent(componentName string, componentReplicas *int32, num
 		return expandMultinodeRoles(componentName, numberOfNodes)
 	case isInterPodGMS:
 		return expandSingleNodeGMSRoles(componentName, component.GetTotalEnginePods())
+	case component.IsGroveScalingGroupForced():
+		return expandSingleNodeScalingGroupRoles(componentName)
 	default:
 		return expandSingleNodeRoles(componentName, componentReplicas)
 	}
@@ -1233,6 +1237,12 @@ func expandSingleNodeRoles(componentName string, componentReplicas *int32) []Ser
 	}
 }
 
+func expandSingleNodeScalingGroupRoles(componentName string) []ServiceRole {
+	return []ServiceRole{
+		{Name: componentName, Role: RoleMain, Replicas: 1},
+	}
+}
+
 func expandMultinodeRoles(componentName string, numberOfNodes int32) []ServiceRole {
 	return []ServiceRole{
 		{Name: componentName + "-" + commonconsts.GroveRoleSuffixLeader, Role: RoleLeader, Replicas: 1},
@@ -1274,7 +1284,7 @@ func LongestPodCliqueNameForDGDComponent(
 	component *v1beta1.DynamoComponentDeploymentSharedSpec,
 ) string {
 	lowerComponentName := strings.ToLower(componentName)
-	if component == nil || (component.GetNumberOfNodes() <= 1 && !component.IsInterPodGMSEnabled()) {
+	if component == nil || !component.UsesPCSG() {
 		return lowerComponentName
 	}
 
@@ -1301,7 +1311,7 @@ func PCSNameForDGD(dgdName string, components []v1beta1.DynamoComponentDeploymen
 		componentName := component.ComponentName
 		lowerName := strings.ToLower(componentName)
 		var budget int
-		if component.GetNumberOfNodes() > 1 || component.IsInterPodGMSEnabled() {
+		if component.UsesPCSG() {
 			// PCSG = lowerName, PCLQ = longest rendered role name.
 			budget = len(lowerName) + len(LongestPodCliqueNameForDGDComponent(componentName, component))
 		} else {
@@ -2537,7 +2547,7 @@ func GenerateGrovePodCliqueSet(
 		isMultinode := numberOfNodes > 1
 		isInterPodGMS := component.IsInterPodGMSEnabled()
 		isInterPodFailover := component.IsInterPodFailoverEnabled()
-		usesPCSG := isMultinode || isInterPodGMS
+		usesPCSG := component.UsesPCSG()
 		roles := expandRolesForComponent(componentName, component.Replicas, numberOfNodes, component)
 		var cliqueNames []string