refactor: move kai-scheduler and DRA driver to base overlay for CNCF AI conformance (#139)

Signed-off-by: Yuan Chen <yuanchen97@gmail.com>

Author: yuanchen8911

recipes/overlays/base.yaml

@@ -69,4 +69,20 @@ spec:
       version: 1.19.2
       valuesFile: components/k8s-ephemeral-storage-metrics/values.yaml
       dependencyRefs:
-        - kube-prometheus-stack
+        - kube-prometheus-stack
+
+    - name: nvidia-dra-driver-gpu
+      type: Helm
+      source: https://helm.ngc.nvidia.com/nvidia
+      version: "25.12.0"
+      valuesFile: components/nvidia-dra-driver-gpu/values.yaml
+      dependencyRefs:
+        - gpu-operator
+
+    - name: kai-scheduler
+      type: Helm
+      source: oci://ghcr.io/nvidia/kai-scheduler
+      version: v0.12.14
+      valuesFile: components/kai-scheduler/values.yaml
+      dependencyRefs:
+        - gpu-operator

recipes/overlays/h100-eks-ubuntu-inference-dynamo.yaml

@@ -75,11 +75,6 @@ spec:
 
     - name: nvidia-dra-driver-gpu
       type: Helm
-      source: https://helm.ngc.nvidia.com/nvidia
-      version: "25.12.0"
-      valuesFile: components/nvidia-dra-driver-gpu/values.yaml
-      dependencyRefs:
-        - gpu-operator
       overrides:
         gpuResourcesEnabledOverride: true
         # EKS has no control-plane nodes — remove the default nodeAffinity
@@ -89,14 +84,6 @@ spec:
           tolerations:
             - operator: Exists
 
-    - name: kai-scheduler
-      type: Helm
-      source: oci://ghcr.io/nvidia/kai-scheduler
-      version: v0.12.14
-      valuesFile: components/kai-scheduler/values.yaml
-      dependencyRefs:
-        - gpu-operator
-
     - name: dynamo-crds
       type: Helm
       source: https://helm.ngc.nvidia.com/nvidia/ai-dynamo

recipes/overlays/h100-kind-inference-dynamo.yaml

@@ -34,14 +34,6 @@ spec:
       value: ">= 1.34"
 
   componentRefs:
-    - name: kai-scheduler
-      type: Helm
-      source: oci://ghcr.io/nvidia/kai-scheduler
-      version: v0.12.14
-      valuesFile: components/kai-scheduler/values.yaml
-      dependencyRefs:
-        - gpu-operator
-
     - name: dynamo-crds
       type: Helm
       source: https://helm.ngc.nvidia.com/nvidia/ai-dynamo

recipes/overlays/kind.yaml

@@ -179,15 +179,9 @@ spec:
                 cpu: 500m
                 memory: 1Gi
 
-    # nvidia-dra-driver-gpu: DRA driver for GPU resource management
-    # For kind with GPU passthrough, use "/" as driver root (host drivers)
+    # nvidia-dra-driver-gpu: override driver root for kind GPU passthrough
     - name: nvidia-dra-driver-gpu
       type: Helm
-      source: https://helm.ngc.nvidia.com/nvidia
-      version: "25.12.0"
-      valuesFile: components/nvidia-dra-driver-gpu/values.yaml
-      dependencyRefs:
-        - gpu-operator
       overrides:
         # Use "/" for host-installed drivers (kind GPU passthrough)
         nvidiaDriverRoot: "/"

tests/chainsaw/cli/cuj1-training/assert-recipe.yaml

@@ -40,8 +40,10 @@ componentRefs: ## alphabetically sorted
   - name: cert-manager
   - name: gpu-operator
   - name: k8s-ephemeral-storage-metrics
+  - name: kai-scheduler
   - name: kube-prometheus-stack
   - name: kubeflow-trainer
+  - name: nvidia-dra-driver-gpu
   - name: nvsentinel
   - name: prometheus-adapter
   - name: skyhook-customizations
@@ -51,9 +53,11 @@ deploymentOrder:
   - aws-efa
   - cert-manager
   - gpu-operator
+  - kai-scheduler
   - kube-prometheus-stack
   - k8s-ephemeral-storage-metrics
   - kubeflow-trainer
+  - nvidia-dra-driver-gpu
   - nvsentinel
   - prometheus-adapter
   - skyhook-operator