llm-slo-decision-gen: schedule onto tainted nodes; slo-api 0.2.1

charts/llm-slo-decision-gen/Chart.yaml

@@ -17,7 +17,7 @@ description: |
   separate chart (llmscaleoperator / llmscaleoperator-system). Deploy
   that one first if you need autoscaling, not just SLO storage.
 type: application
-version: 0.3.5
+version: 0.3.6
 # Matches decisionGen.image.tag (4pdosc/llm-scaler-decision-gen:0.7.0), the
 # first release that reads LLMSLORequirement from inference.modelsphere.dev.
 appVersion: "0.7.0"

charts/llm-slo-decision-gen/templates/decision-gen-deployment.yaml

@@ -24,6 +24,18 @@ spec:
         app: decision-gen
     spec:
       serviceAccountName: decision-gen
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
       containers:
         - name: decision-gen
           image: "{{ .Values.decisionGen.image.repository }}:{{ .Values.decisionGen.image.tag }}"

charts/llm-slo-decision-gen/templates/decision-gen-unstable-deployment.yaml

@@ -24,6 +24,18 @@ spec:
         app: decision-gen-unstable
     spec:
       serviceAccountName: decision-gen
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
       containers:
         - name: decision-gen
           # Same image as stable; unstable differs only by env (LOG_LEVEL=DEBUG).

charts/llm-slo-decision-gen/templates/slo-api-deployment.yaml

@@ -24,6 +24,18 @@ spec:
         app: slo-api
     spec:
       serviceAccountName: slo-api
+      {{- with .Values.nodeSelector }}
+      nodeSelector:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.affinity }}
+      affinity:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
+      {{- with .Values.tolerations }}
+      tolerations:
+        {{- toYaml . | nindent 8 }}
+      {{- end }}
       containers:
         - name: slo-api
           image: "{{ .Values.sloApi.image.repository }}:{{ .Values.sloApi.image.tag }}"

charts/llm-slo-decision-gen/values.yaml

@@ -4,6 +4,14 @@
 # `helm upgrade --install -n <ns>`. Workloads use {{ .Release.Namespace }}.
 # Service DNS is http://decision-gen.<Release.Namespace>.svc:80 (and siblings).
 
+# Scheduling for every pod of this release (decision-gen, its unstable twin and slo-api): they are
+# small CPU-only services that land wherever the cluster lets them. Set tolerations when all nodes
+# are tainted -- an NPU cluster can carry huawei.com/npu:NoSchedule on every node, and then nothing
+# here schedules without one.
+tolerations: []
+nodeSelector: {}
+affinity: {}
+
 # ---------------------------------------------------------------------------
 # decision-gen: emits per-service replica decisions on :8080/decisions.
 # The "unstable" variant (when enabled) is a second Deployment/Service so a
@@ -82,7 +90,7 @@ sloApi:
   replicas: 1
   image:
     repository: 4pdosc/llm-slo-api
-    tag: "0.2.0"
+    tag: "0.2.1"
     pullPolicy: IfNotPresent
   service:
     port: 80