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/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