feat: add ComfyUI, rebalance GPU allocation

GPU rebalance (4× V100 32GB on worker-1):
- reasoning: 2 GPU (unchanged, PP=2 for Qwen3-32B)
- ornith: 2→1 GPU (scale to 1 replica, ornith:35b only)
- comfyui: 0→1 GPU (new)
- embeddings/reranker: 0 GPU (CPU, unchanged)

qwen2.5:3b-instruct moved to CPU on talos-cp-2 (144GB RAM).
Separate Ollama deployment + 5Gi PVC, pulls model on first start.

Gateway config updated in homelab-frontend (separate commit).

Co-authored-by: poimen <[email protected]>

k8s/apps/llm-serving/kustomization.yaml

@@ -14,6 +14,7 @@ resources:
   - ornith.yaml
   - reasoning.yaml
   - reranker.yaml
+  - qwen-cpu.yaml
   - networkpolicy.yaml
 # No namespace transformer: every file sets its own, and the transformer would
 # rewrite metadata.namespace on anything cross-namespace added later.

k8s/apps/llm-serving/ornith.yaml

@@ -33,12 +33,8 @@ spec:
 
         ollama pull ornith:35b
 
-        ollama pull qwen2.5:3b-instruct
-
         ollama run ornith:35b "ok" >/dev/null 2>&1 || true
 
-        ollama run qwen2.5:3b-instruct "ok" >/dev/null 2>&1 || true
-
         wait $SERVE_PID
 
         '
@@ -54,7 +50,7 @@ spec:
       - name: OLLAMA_NUM_PARALLEL
         value: '1'
       - name: OLLAMA_MAX_LOADED_MODELS
-        value: '2'
+        value: '1'
       image: ollama/ollama:0.32.9@sha256:1685741456770df6e3cceb2a945a5f75e020f658d1701509668d6f4688f1dd3f
       name: kserve-container
       ports:
@@ -65,8 +61,7 @@ spec:
           command:
           - /bin/sh
           - -c
-          - ollama ps 2>/dev/null | grep -q ornith && ollama ps 2>/dev/null |
-            grep -q qwen2.5
+          - ollama ps 2>/dev/null | grep -q ornith
         periodSeconds: 10
       resources:
         limits:
@@ -82,8 +77,7 @@ spec:
           command:
           - /bin/sh
           - -c
-          - ollama ps 2>/dev/null | grep -q ornith && ollama ps 2>/dev/null |
-            grep -q qwen2.5
+          - ollama ps 2>/dev/null | grep -q ornith
         failureThreshold: 120
         periodSeconds: 15
       volumeMounts:
@@ -91,13 +85,10 @@ spec:
         name: models
     deploymentStrategy:
       type: Recreate
-    # 2 replicas -- each its own GPU, each loading both ornith:35b and
-    # qwen2.5:3b-instruct -- so 2 concurrent implementer-style calls each
-    # get an independent instance instead of contending on one, at the
-    # cost of judge/qwen traffic still sharing whichever replica an
-    # implementer call also lands on.
-    maxReplicas: 2
-    minReplicas: 2
+    # 1 replica -- ornith:35b only. qwen2.5:3b moved to CPU on cp-2.
+    # Frees 1 GPU for ComfyUI.
+    maxReplicas: 1
+    minReplicas: 1
     nodeSelector:
       kubernetes.io/hostname: worker-1
     runtimeClassName: nvidia

k8s/apps/llm-serving/qwen-cpu.yaml

@@ -0,0 +1,115 @@
+# qwen2.5:3b-instruct on CPU (talos-cp-2, 144GB RAM, 24 cores).
+# Moved off GPU to free a V100 for ComfyUI. Latency ~10x slower
+# than GPU but sufficient for lightweight tasks (summarization,
+# classification, quick answers).
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: qwen-cpu
+  namespace: llm-serving
+  labels:
+    app: qwen-cpu
+    app.kubernetes.io/name: qwen-cpu
+    app.kubernetes.io/part-of: llm-serving
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: qwen-cpu
+  template:
+    metadata:
+      labels:
+        app: qwen-cpu
+        app.kubernetes.io/name: qwen-cpu
+        app.kubernetes.io/part-of: llm-serving
+    spec:
+      nodeSelector:
+        kubernetes.io/hostname: talos-cp-2
+      tolerations:
+        - key: node-role.kubernetes.io/control-plane
+          operator: Exists
+          effect: NoSchedule
+      containers:
+        - name: ollama
+          image: ollama/ollama:0.32.9@sha256:1685741456770df6e3cceb2a945a5f75e020f658d1701509668d6f4688f1dd3f
+          command: ["/bin/sh", "-c"]
+          args:
+            - |
+              ollama serve &
+              SERVE_PID=$!
+              until ollama list >/dev/null 2>&1; do sleep 2; done
+              ollama pull qwen2.5:3b-instruct
+              ollama run qwen2.5:3b-instruct "ok" >/dev/null 2>&1 || true
+              wait $SERVE_PID
+          env:
+            - name: OLLAMA_HOST
+              value: "0.0.0.0:8080"
+            - name: OLLAMA_MODELS
+              value: /root/.ollama/models
+            - name: OLLAMA_CONTEXT_LENGTH
+              value: "32768"
+            - name: OLLAMA_KEEP_ALIVE
+              value: "-1"
+            - name: OLLAMA_MAX_LOADED_MODELS
+              value: "1"
+            - name: OLLAMA_NUM_PARALLEL
+              value: "2"
+          ports:
+            - containerPort: 8080
+              protocol: TCP
+          readinessProbe:
+            exec:
+              command: ["/bin/sh", "-c", "ollama ps 2>/dev/null | grep -q qwen2.5"]
+            periodSeconds: 10
+          startupProbe:
+            exec:
+              command: ["/bin/sh", "-c", "ollama ps 2>/dev/null | grep -q qwen2.5"]
+            failureThreshold: 60
+            periodSeconds: 10
+          resources:
+            requests:
+              cpu: "4"
+              memory: 4Gi
+            limits:
+              cpu: "8"
+              memory: 8Gi
+          volumeMounts:
+            - mountPath: /root/.ollama
+              name: ollama-data
+      volumes:
+        - name: ollama-data
+          persistentVolumeClaim:
+            claimName: qwen-cpu-data
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: qwen-cpu
+  namespace: llm-serving
+  labels:
+    app: qwen-cpu
+    app.kubernetes.io/part-of: llm-serving
+spec:
+  selector:
+    app: qwen-cpu
+  ports:
+    - port: 80
+      targetPort: 8080
+      protocol: TCP
+---
+# Small PVC for qwen2.5:3b model weights (~1.9GB).
+# Separate from llm-models PVC which is pinned to worker-1.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: qwen-cpu-data
+  namespace: llm-serving
+spec:
+  accessModes:
+    - ReadWriteOnce
+  storageClassName: longhorn
+  resources:
+    requests:
+      storage: 5Gi