fix(binder): allow explicit CDI conf override (#951)

Author: enoodle

CHANGELOG.md

@@ -6,6 +6,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [v0.12.11] - 2026-02-03
+### Fixed
+- Added `binder.cdiEnabled` Helm value to allow explicit override of CDI auto-detection for environments without ClusterPolicy fixing compatibility issues in Openshift
+
 ## [v0.12.9] - 2026-01-21
 
 ### Fixed

README.md

@@ -67,6 +67,7 @@ When `gpu-operator` <v25.10.0 is installed, the following flag should be added t
 ```
 --set admission.gpuPodRuntimeClassName=null
 ```
+If CDI is enabled also add `--set binder.cdiEnabled=true`.
 
 ## Support & Breaking changes
 For details on our release lifecycle, LTS versions, and supported releases, see the [Support Policy](SUPPORT.md).

deployments/kai-scheduler/crds/kai.scheduler_configs.yaml

@@ -1116,6 +1116,11 @@ spec:
               binder:
                 description: Binder specifies configuration for the binder
                 properties:
+                  cdiEnabled:
+                    description: |-
+                      CDIEnabled Specifies if the gpu device plugin uses the cdi devices api to set gpu devices to the pods
+                      leave empty if unsure to let the operator auto detect using ClusterPolicy (nvidia gpu-operator only)
+                    type: boolean
                   maxConcurrentReconciles:
                     description: MaxConcurrentReconciles is the maximum number of
                       concurrent reconciles for both pods and BindRequests

deployments/kai-scheduler/templates/kai-config.yaml

@@ -67,6 +67,9 @@ spec:
       podResources:
         {{- toYaml .Values.binder.resourceReservationPodResources | nindent 8 }}
       {{- end }}
+    {{- if hasKey .Values.binder "cdiEnabled" }}
+    cdiEnabled: {{ .Values.binder.cdiEnabled }}
+    {{- end }}
 
   podGroupController:
     service:

pkg/apis/kai/v1/binder/binder.go

@@ -42,6 +42,11 @@ type Binder struct {
 
 	// MetricsPort specifies the metrics service port
 	MetricsPort *int `json:"metricsPort,omitempty"`
+
+	// CDIEnabled Specifies if the gpu device plugin uses the cdi devices api to set gpu devices to the pods
+	// leave empty if unsure to let the operator auto detect using ClusterPolicy (nvidia gpu-operator only)
+	// +kubebuilder:validation:Optional
+	CDIEnabled *bool `json:"cdiEnabled,omitempty"`
 }
 
 func (b *Binder) SetDefaultsWhereNeeded(replicaCount *int32) {
@@ -76,7 +81,6 @@ func (b *Binder) SetDefaultsWhereNeeded(replicaCount *int32) {
 
 	b.ProbePort = common.SetDefault(b.ProbePort, ptr.To(8081))
 	b.MetricsPort = common.SetDefault(b.MetricsPort, ptr.To(8080))
-
 }
 
 type ResourceReservation struct {

pkg/apis/kai/v1/binder/zz_generated.deepcopy.go

@@ -164,6 +164,35 @@ var _ = Describe("Binder", func() {
 					Expect(deploymentT).NotTo(BeNil())
 					Expect((*deploymentT).Spec.Template.Spec.Containers[0].Args).To(ContainElement("--cdi-enabled=false"))
 				})
+
+				It("uses explicit CDIEnabled=true from config, ignoring ClusterPolicy", func(ctx context.Context) {
+					clusterPolicy.Spec.CDI.Default = ptr.To(false)
+					Expect(fakeKubeClient.Create(ctx, clusterPolicy)).To(Succeed())
+
+					kaiConfig.Spec.Binder.CDIEnabled = ptr.To(true)
+
+					objects, err := b.DesiredState(ctx, fakeKubeClient, kaiConfig)
+					Expect(err).To(BeNil())
+
+					deploymentT := test_utils.FindTypeInObjects[*appsv1.Deployment](objects)
+					Expect(deploymentT).NotTo(BeNil())
+					Expect((*deploymentT).Spec.Template.Spec.Containers[0].Args).To(ContainElement("--cdi-enabled=true"))
+				})
+
+				It("uses explicit CDIEnabled=false from config, ignoring ClusterPolicy", func(ctx context.Context) {
+					clusterPolicy.Spec.CDI.Enabled = ptr.To(true)
+					clusterPolicy.Spec.CDI.Default = ptr.To(true)
+					Expect(fakeKubeClient.Create(ctx, clusterPolicy)).To(Succeed())
+
+					kaiConfig.Spec.Binder.CDIEnabled = ptr.To(false)
+
+					objects, err := b.DesiredState(ctx, fakeKubeClient, kaiConfig)
+					Expect(err).To(BeNil())
+
+					deploymentT := test_utils.FindTypeInObjects[*appsv1.Deployment](objects)
+					Expect(deploymentT).NotTo(BeNil())
+					Expect((*deploymentT).Spec.Template.Spec.Containers[0].Args).To(ContainElement("--cdi-enabled=false"))
+				})
 			})
 		})
 

pkg/operator/operands/binder/binder_test.go

@@ -48,9 +48,14 @@ func (b *Binder) deploymentForKAIConfig(
 		return nil, err
 	}
 
-	cdiEnabled, err := isCdiEnabled(ctx, runtimeClient)
-	if err != nil {
-		return nil, err
+	var cdiEnabled bool
+	if config.CDIEnabled != nil {
+		cdiEnabled = *config.CDIEnabled
+	} else {
+		cdiEnabled, err = isCdiEnabled(ctx, runtimeClient)
+		if err != nil {
+			return nil, err
+		}
 	}
 
 	deployment.Spec.Strategy.Type = appsv1.RecreateDeploymentStrategyType