feat(modeldeployment): add nodeLabelSelector for InferenceSet GPU node selection

charts/modeldeployment/README.md

@@ -41,7 +41,8 @@ over plaintext gRPC and **no `DestinationRule` is required**.
 | `model`                   | required | `""`                                                                 | Preset model name. Used only as `spec.template.inference.preset.name`.                     |
 | `config`                  | optional | `""`                                                                 | Name of a ConfigMap in the deployment namespace containing custom inference arguments. Rendered as `spec.template.inference.config`. |
 | `modelAccessSecret`       | optional | `""`                                                                 | Name of a Secret containing Hugging Face model access credentials. Rendered as `spec.template.inference.preset.presetOptions.modelAccessSecret`. |
-| `instanceType`            | optional | `Standard_NV36ads_A10_v5`                                            | VM instance type for the underlying nodes. Set `""` to omit `spec.template.resource.instanceType`, as required on BYO-node clusters (KAITO `nodeProvisioner=byo`). |
+| `instanceType`            | optional | `Standard_NV36ads_A10_v5`                                            | VM instance type for the underlying nodes. Set `""` to omit `spec.template.resource.instanceType`. Required with node auto-provisioning; optional on BYO-node clusters (KAITO `nodeProvisioner=byo`), where it restricts GPU nodes to `node.kubernetes.io/instance-type=<instanceType>`. |
+| `nodeLabelSelector`       | optional | `{}`                                                                 | GPU node [LabelSelector](https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/#resources-that-support-set-based-requirements) (`matchLabels` / `matchExpressions`), rendered as `spec.template.resource.labelSelector`. Empty omits it: with node auto-provisioning it inherits `spec.labelSelector`; on BYO-node clusters GPU nodes are auto-selected. Requires a KAITO release with kaito-project/kaito#2444 (v0.13.0+); older CRDs silently drop the field. |
 | `replicas`                | optional | _empty_                                                              | InferenceSet replicas (`spec.replicas`). Empty leaves `spec.replicas` unset (CRD default `1`; with `enableScaling=true`, KEDA moves it into `[minReplicas, maxReplicas]`). When set with `enableScaling=true` it must lie within `[minReplicas, maxReplicas]` and is re-applied only when its value changes (see [Replica ownership](#replica-ownership)). |
 | `minReplicas`             | required when `enableScaling=true` | _empty_                                    | Autoscaling lower bound, wired to `scaledobject.kaito.sh/min-replicas`. Ignored when `enableScaling=false`. `0` enables [scale to zero](#scale-to-zero). |
 | `enableScaling`           | optional | `false`                                                              | Wired to `scaledobject.kaito.sh/auto-provision`. Gates `minReplicas`, `maxReplicas` and the entire `scaling` block. |

charts/modeldeployment/templates/inferenceset.yaml

@@ -233,12 +233,20 @@ spec:
         {{- end }}
       {{- end }}
     {{- /*
-    Empty instanceType omits the field, which KAITO requires on BYO
-    (nodeProvisioner=byo) clusters where nodes are not auto-provisioned.
+    Empty instanceType / nodeLabelSelector omit their fields. On BYO
+    (nodeProvisioner=byo) clusters both are optional node filters and KAITO
+    auto-selects GPU nodes when both are empty; with node auto-provisioning an
+    omitted labelSelector inherits spec.labelSelector.
     */}}
     resource:
+      {{- if or .Values.instanceType .Values.nodeLabelSelector }}
       {{- with .Values.instanceType }}
       instanceType: {{ . | quote }}
+      {{- end }}
+      {{- with .Values.nodeLabelSelector }}
+      labelSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
       {{- else }} {}
       {{- end }}
     inference:

charts/modeldeployment/values.schema.json

@@ -85,7 +85,36 @@
     },
     "instanceType": {
       "type": "string",
-      "description": "VM instance type used for the underlying GPU nodes. Empty omits spec.template.resource.instanceType, as required on BYO-node clusters (KAITO nodeProvisioner=byo)."
+      "description": "VM instance type used for the underlying GPU nodes. Empty omits spec.template.resource.instanceType. Required with node auto-provisioning; optional on BYO-node clusters (KAITO nodeProvisioner=byo), where it restricts GPU nodes to node.kubernetes.io/instance-type=<instanceType>."
+    },
+    "nodeLabelSelector": {
+      "description": "LabelSelector for the GPU nodes, rendered as spec.template.resource.labelSelector. Empty omits it: with node auto-provisioning it then inherits spec.labelSelector; on BYO-node clusters GPU nodes are auto-selected. Requires KAITO v0.13.0+.",
+      "anyOf": [
+        {"type": "null"},
+        {
+          "type": "object",
+          "additionalProperties": false,
+          "properties": {
+            "matchLabels": {
+              "type": "object",
+              "additionalProperties": {"type": "string"}
+            },
+            "matchExpressions": {
+              "type": "array",
+              "items": {
+                "type": "object",
+                "additionalProperties": false,
+                "required": ["key", "operator"],
+                "properties": {
+                  "key": {"type": "string", "minLength": 1},
+                  "operator": {"type": "string", "enum": ["In", "NotIn", "Exists", "DoesNotExist"]},
+                  "values": {"type": "array", "items": {"type": "string"}}
+                }
+              }
+            }
+          }
+        }
+      ]
     },
     "replicas": {
       "anyOf": [

charts/modeldeployment/values.yaml

@@ -16,9 +16,16 @@ config: ""
 # modelAccessSecret names the Secret holding the Hugging Face access credentials.
 modelAccessSecret: ""
 
-# instanceType is the VM instance type of the GPU nodes; set "" on BYO-node clusters to omit it.
+# instanceType is the VM instance type of the GPU nodes; "" omits it. Required with node auto-provisioning;
+# optional on BYO-node clusters, where it restricts GPU nodes to node.kubernetes.io/instance-type=<instanceType>.
 instanceType: "Standard_NV36ads_A10_v5"
 
+# nodeLabelSelector is a LabelSelector (matchLabels / matchExpressions) for the GPU nodes, rendered as
+# spec.template.resource.labelSelector; empty omits it. With node auto-provisioning an empty value inherits the
+# InferenceSet spec.labelSelector; on BYO-node clusters GPU nodes are auto-selected. Requires a KAITO release with
+# kaito-project/kaito#2444 (v0.13.0+); older CRDs silently drop it.
+nodeLabelSelector: {}
+
 # replicas sets spec.replicas when not empty (otherwise the CRD default of 1); with enableScaling it must lie
 # within [minReplicas, maxReplicas] and is reapplied on upgrade only when its value changes.
 replicas: ""