feat(llm): Ollama warmup bei Start - Modelle permanent im VRAM

- warmup_ollama() laedt qwen3:30b-a3b + nomic-embed-text mit keep_alive=-1
- Wird beim Bot-Start in post_init() aufgerufen (via asyncio.to_thread)
- keep_alive=-1 nur ueber native Ollama API (/api/generate) moeglich
- GPU haelt 22.6/24 GB permanent: Text + Embeddings ohne Swap

homelab-ai-bot/llm.py

@@ -25,6 +25,20 @@ FALLBACK_MODEL = "qwen2.5:14b"
 MAX_TOOL_ROUNDS = 3
 OLLAMA_MODELS = {MODEL_LOCAL, FALLBACK_MODEL}
 
+
+def warmup_ollama():
+    """Laedt Hauptmodell + Embedding permanent in VRAM (keep_alive=-1)."""
+    for model in [MODEL_LOCAL, "nomic-embed-text"]:
+        try:
+            requests.post(
+                f"{OLLAMA_BASE}/api/generate",
+                json={"model": model, "prompt": "", "keep_alive": -1},
+                timeout=120,
+            )
+            log.info("Ollama warmup: %s permanent geladen", model)
+        except Exception as e:
+            log.warning("Ollama warmup fehlgeschlagen fuer %s: %s", model, e)
+
 PASSTHROUGH_TOOLS = {"get_temperaturen", "get_energie", "get_heizung"}
 
 _LOCAL_OVERRIDES = [
@@ -269,7 +283,6 @@ def _call_api(messages: list, api_key: str, use_tools: bool = True,
         headers = {"Content-Type": "application/json"}
         timeout = _ollama_timeout_for(chosen)
         _add_no_think(payload.get("messages", []))
-        payload["keep_alive"] = -1
     else:
         url = f"{OPENROUTER_BASE}/chat/completions"
         headers = {"Authorization": f"Bearer {api_key}"}

homelab-ai-bot/telegram_bot.py

@@ -1025,6 +1025,7 @@ def main():
         await application.bot.set_my_commands(BOT_COMMANDS)
         log.info("Kommandomenü registriert")
         asyncio.create_task(_watchdog_loop())
+        asyncio.create_task(asyncio.to_thread(llm.warmup_ollama))
         asyncio.create_task(_monitor_loop(application))
         log.info("Monitor-Loop aktiv (alle 10 Min)")
         if application.job_queue is None: