feat: add ComfyUI + rebalance GPU allocation (#17)

## GPU Rebalance (4× V100 32GB)

| Pod | Before | After |
|-----|--------|-------|
| reasoning (PP=2) | 2 GPU | 2 GPU |
| ornith | 2 GPU (2 replicas) | 1 GPU (1 replica) |
| comfyui | — | 1 GPU (**new**) |
| qwen-cpu | — | CPU on cp-2 (**new**) |
| embeddings/reranker | CPU | CPU |

## Changes

- `ornith.yaml`: scale 2→1, remove qwen2.5 co-loading, MAX_LOADED_MODELS=1
- `qwen-cpu.yaml`: new Ollama deployment on talos-cp-2 (144GB RAM), 5Gi PVC
- `k8s/apps/comfyui/`: new ComfyUI deployment (1 GPU, 50Gi model PVC, ingress)
- `58-comfyui.yaml`: ArgoCD Application (wave 8)

Gateway route update in separate PR (homelab-frontend).Reviewed-on: #17

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

k8s/apps/comfyui/deployment.yaml

@@ -0,0 +1,80 @@
+# ComfyUI — GPU-accelerated image generation on worker-1.
+# Uses 1x V100 32GB (sm70). Freed by scaling ornith 2→1.
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: comfyui
+  namespace: comfyui
+  labels:
+    app: comfyui
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: comfyui
+  template:
+    metadata:
+      labels:
+        app: comfyui
+    spec:
+      nodeSelector:
+        kubernetes.io/hostname: worker-1
+      runtimeClassName: nvidia
+      containers:
+        - name: comfyui
+          image: ghcr.io/ai-dock/comfyui:v2-cuda-12.1.1-base-22.04
+          ports:
+            - containerPort: 8188
+              protocol: TCP
+          env:
+            - name: NVIDIA_VISIBLE_DEVICES
+              value: "all"
+            - name: COMFYUI_FLAGS
+              value: "--listen 0.0.0.0 --port 8188"
+          resources:
+            requests:
+              cpu: "4"
+              memory: 8Gi
+              nvidia.com/gpu: "1"
+            limits:
+              cpu: "8"
+              memory: 16Gi
+              nvidia.com/gpu: "1"
+          volumeMounts:
+            - mountPath: /workspace/ComfyUI/models
+              name: models
+            - mountPath: /workspace/ComfyUI/output
+              name: output
+          readinessProbe:
+            httpGet:
+              path: /
+              port: 8188
+            periodSeconds: 10
+            initialDelaySeconds: 30
+          startupProbe:
+            httpGet:
+              path: /
+              port: 8188
+            failureThreshold: 60
+            periodSeconds: 10
+      volumes:
+        - name: models
+          persistentVolumeClaim:
+            claimName: comfyui-models
+        - name: output
+          emptyDir: {}
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: comfyui-models
+  namespace: comfyui
+spec:
+  accessModes:
+    - ReadWriteOnce
+  storageClassName: longhorn
+  resources:
+    requests:
+      storage: 50Gi

k8s/apps/comfyui/ingress.yaml

@@ -0,0 +1,25 @@
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: comfyui
+  namespace: comfyui
+  annotations:
+    nginx.ingress.kubernetes.io/proxy-read-timeout: "600"
+    nginx.ingress.kubernetes.io/proxy-send-timeout: "600"
+    nginx.ingress.kubernetes.io/proxy-body-size: "0"
+    # WebSocket support for ComfyUI's live preview
+    nginx.ingress.kubernetes.io/proxy-http-version: "1.1"
+    nginx.ingress.kubernetes.io/proxy-set-headers: "Upgrade"
+spec:
+  ingressClassName: nginx
+  rules:
+    - host: comfy.riotpiao.com
+      http:
+        paths:
+          - path: /
+            pathType: Prefix
+            backend:
+              service:
+                name: comfyui
+                port:
+                  number: 80

k8s/apps/comfyui/kustomization.yaml

@@ -0,0 +1,7 @@
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+
+resources:
+  - deployment.yaml
+  - service.yaml
+  - ingress.yaml

k8s/apps/comfyui/service.yaml

@@ -0,0 +1,14 @@
+apiVersion: v1
+kind: Service
+metadata:
+  name: comfyui
+  namespace: comfyui
+  labels:
+    app: comfyui
+spec:
+  selector:
+    app: comfyui
+  ports:
+    - port: 80
+      targetPort: 8188
+      protocol: TCP

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 |
+          - ollama ps 2>/dev/null | grep -q ornith
-            grep -q qwen2.5
         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 |
+          - ollama ps 2>/dev/null | grep -q ornith
-            grep -q qwen2.5
         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
+    # 1 replica -- ornith:35b only. qwen2.5:3b moved to CPU on cp-2.
-    # qwen2.5:3b-instruct -- so 2 concurrent implementer-style calls each
+    # Frees 1 GPU for ComfyUI.
-    # get an independent instance instead of contending on one, at the
+    maxReplicas: 1
-    # cost of judge/qwen traffic still sharing whichever replica an
+    minReplicas: 1
-    # implementer call also lands on.
-    maxReplicas: 2
-    minReplicas: 2
     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

k8s/argocd/apps/58-comfyui.yaml

@@ -0,0 +1,32 @@
+apiVersion: argoproj.io/v1alpha1
+kind: Application
+metadata:
+  name: comfyui
+  namespace: argocd
+  labels:
+    app.kubernetes.io/name: comfyui
+    app.kubernetes.io/component: image-generation
+  annotations:
+    argocd.argoproj.io/sync-wave: "8"
+spec:
+  project: homelab
+  revisionHistoryLimit: 3
+  source:
+    repoURL: https://forgejo.riotpiao.com/rock/homelab.git
+    targetRevision: main
+    path: k8s/apps/comfyui
+  destination:
+    server: https://kubernetes.default.svc
+    namespace: comfyui
+  syncPolicy:
+    automated:
+      prune: true
+      selfHeal: true
+    syncOptions:
+      - CreateNamespace=true
+    retry:
+      limit: 5
+      backoff:
+        duration: 5s
+        factor: 2
+        maxDuration: 3m