feat: upgrade Ollama to 0.17.0 and add model swap script for qwen3:4b

Bumps Ollama image from 0.7.0 to 0.17.0 and adds update-model.sh to
streamline swapping models (pull new, verify GPU, remove old). Updates
docs to reflect current deployment name and qwen3:4b as the default model.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: web-flow

gitops/clusters/homelab/apps/ollama/deployment.yaml

@@ -20,7 +20,7 @@ spec:
         nvidia.com/gpu.present: "true"
       containers:
         - name: ollama
-          image: ollama/ollama:0.7.0
+          image: ollama/ollama:0.17.0
           imagePullPolicy: IfNotPresent
           args:
             - serve

proxmox/guides/ollama-gpu-server.md

@@ -16,17 +16,27 @@ the models to survive pod restarts.
 1. Pull a model into the running pod:
 
    ```bash
-   kubectl exec deployment/ollama-server -- ollama pull llama3
+   kubectl exec deployment/ollama-gpu -- ollama pull qwen3:4b
    ```
 
 2. Verify the model is available:
 
    ```bash
-   kubectl exec deployment/ollama-server -- ollama list
+   kubectl exec deployment/ollama-gpu -- ollama list
    ```
 
 3. Send a test request to the service:
 
    ```bash
-   curl http://<service-ip>:80/api/generate -d '{"model":"llama3","prompt":"Hello"}'
+   curl http://<service-ip>:80/api/generate -d '{"model":"qwen3:4b","prompt":"Hello"}'
    ```
+
+### Model Swap Script
+
+To swap models (pull new, remove old, verify GPU):
+
+```bash
+scripts/ollama/update-model.sh qwen3:4b qwen2.5:3b
+```
+
+Add `--ha` to also update the Home Assistant conversation agent.

scripts/ollama/update-model.sh

@@ -0,0 +1,126 @@
+#!/bin/bash
+# Swap Ollama model: pull new model, remove old, verify GPU loading
+# Usage: update-model.sh [new_model] [old_model]
+#
+# Examples:
+#   update-model.sh                            # default: pull qwen3:4b, remove qwen2.5:3b
+#   update-model.sh qwen3:4b                   # pull qwen3:4b only (no removal)
+#   update-model.sh qwen3:4b qwen2.5:3b        # pull new, remove old
+#
+# This script:
+#   1. Pulls the new model via Ollama API
+#   2. Verifies the new model loaded on GPU
+#   3. Removes the old model to free PVC space (if specified)
+#   4. Updates HA conversation agent (optional, with --ha flag)
+
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+KUBECONFIG="${KUBECONFIG:-$HOME/kubeconfig}"
+export KUBECONFIG
+
+OLLAMA_NS="${OLLAMA_NS:-ollama}"
+NEW_MODEL="${1:-qwen3:4b}"
+OLD_MODEL="${2:-qwen2.5:3b}"
+UPDATE_HA=false
+
+# Parse flags
+for arg in "$@"; do
+    [[ "$arg" == "--ha" ]] && UPDATE_HA=true
+done
+
+# Get pod name
+POD=$(kubectl get pods -n "$OLLAMA_NS" -l app=ollama-gpu -o jsonpath='{.items[0].metadata.name}' 2>/dev/null)
+if [[ -z "$POD" ]]; then
+    echo "ERROR: No ollama-gpu pod found in namespace $OLLAMA_NS"
+    exit 1
+fi
+
+# Get Ollama URL
+OLLAMA_IP=$(kubectl get svc -n "$OLLAMA_NS" -o jsonpath='{.items[0].status.loadBalancer.ingress[0].ip}' 2>/dev/null)
+OLLAMA_PORT=$(kubectl get svc -n "$OLLAMA_NS" -o jsonpath='{.items[0].spec.ports[0].port}' 2>/dev/null)
+OLLAMA_URL="http://${OLLAMA_IP}:${OLLAMA_PORT:-80}"
+
+echo "=== Ollama Model Swap ==="
+echo "Pod:       $POD"
+echo "URL:       $OLLAMA_URL"
+echo "New model: $NEW_MODEL"
+echo "Old model: $OLD_MODEL"
+echo ""
+
+# 1. Show current models
+echo "--- Current Models ---"
+kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama list 2>/dev/null || echo "  No models found"
+echo ""
+
+# 2. Pull new model
+echo "--- Pulling $NEW_MODEL ---"
+echo "This may take a few minutes depending on model size..."
+kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama pull "$NEW_MODEL"
+echo ""
+
+# 3. Verify new model exists
+echo "--- Verifying $NEW_MODEL ---"
+if kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama list 2>/dev/null | grep -q "$NEW_MODEL"; then
+    echo "OK: $NEW_MODEL is available"
+else
+    echo "ERROR: $NEW_MODEL not found after pull"
+    exit 1
+fi
+echo ""
+
+# 4. Test inference on GPU
+echo "--- Testing inference (GPU check) ---"
+RESPONSE=$(curl -s --connect-timeout 30 --max-time 120 \
+    -X POST "$OLLAMA_URL/api/generate" \
+    -d "{\"model\":\"$NEW_MODEL\",\"prompt\":\"Say hello in one sentence.\",\"stream\":false}" 2>/dev/null)
+
+echo "$RESPONSE" | python3 -c "
+import json, sys
+try:
+    d = json.load(sys.stdin)
+    resp = d.get('response', 'N/A')
+    eval_s = d.get('eval_duration', 0) / 1e9
+    tokens = d.get('eval_count', 0)
+    tps = tokens / eval_s if eval_s > 0 else 0
+    print(f'  Response: {resp[:100]}')
+    print(f'  Speed:    {tps:.1f} tokens/sec ({tokens} tokens in {eval_s:.1f}s)')
+    if tps < 5:
+        print(f'  WARNING: Likely running on CPU!')
+    else:
+        print(f'  OK: GPU acceleration confirmed')
+except Exception as e:
+    print(f'ERROR: {e}')
+    sys.exit(1)
+" 2>/dev/null
+echo ""
+
+# 5. GPU memory check
+echo "--- GPU Memory ---"
+kubectl exec -n "$OLLAMA_NS" "$POD" -- nvidia-smi --query-gpu=memory.used,memory.total --format=csv,noheader 2>/dev/null || echo "  nvidia-smi unavailable"
+echo ""
+
+# 6. Remove old model (if different from new)
+if [[ -n "$OLD_MODEL" && "$OLD_MODEL" != "$NEW_MODEL" ]]; then
+    if kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama list 2>/dev/null | grep -q "$OLD_MODEL"; then
+        echo "--- Removing old model: $OLD_MODEL ---"
+        kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama rm "$OLD_MODEL"
+        echo "Removed $OLD_MODEL"
+    else
+        echo "--- Old model $OLD_MODEL not present (skipping removal) ---"
+    fi
+    echo ""
+fi
+
+# 7. Final model list
+echo "--- Final Models ---"
+kubectl exec -n "$OLLAMA_NS" "$POD" -- ollama list 2>/dev/null
+echo ""
+
+# 8. Optionally update HA
+if [[ "$UPDATE_HA" == true ]]; then
+    echo "--- Updating HA conversation agent ---"
+    "$SCRIPT_DIR/set-ha-model.sh" "$NEW_MODEL"
+fi
+
+echo "=== Done ==="