diff --git a/BEDIENUNGSANLEITUNG.md b/BEDIENUNGSANLEITUNG.md
index e13a294..ccfef77 100644
--- a/BEDIENUNGSANLEITUNG.md
+++ b/BEDIENUNGSANLEITUNG.md
@@ -1755,7 +1755,10 @@ Kategorie, Textausschnitt).
#### Status
Zeigt aktives Profil, Provider-Konfiguration, Laufzeit-Metriken und verfügbare Provider.
-Am Ende: **⬇ voice-assistant.db herunterladen** — lädt die SQLite-Datenbank als Backup.
+Zusätzlich eine **LLM-Backend-Karte** (read-only): aktives Backend (Ollama/llama.cpp),
+Modell, ob die Backends laufen, geladene Ollama-Modelle und die **GPU-Auslastung** je
+Karte als Balken. Am Ende: **⬇ voice-assistant.db herunterladen** — lädt die
+SQLite-Datenbank als Backup.
#### Metriken
diff --git a/app/admin_llm.py b/app/admin_llm.py
new file mode 100644
index 0000000..8a5acf1
--- /dev/null
+++ b/app/admin_llm.py
@@ -0,0 +1,115 @@
+"""Read-only Statusabfragen rund um das lokale LLM-Backend (fuer das Admin-Panel).
+
+Alles nur lesend, ohne sudo: `docker ps`, `ollama ps`, `nvidia-smi`, `systemctl --user`.
+Fehlende Tools oder Fehler fuehren zu sicheren Defaults (None/[]), nie zu Exceptions.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import os
+import shutil
+
+from app.config import settings
+
+
+async def _run(cmd: list[str], timeout: float = 6.0) -> str | None:
+ """Fuehrt ein Kommando aus (shell=False) und liefert stdout, oder None bei Fehler."""
+ if not shutil.which(cmd[0]):
+ return None
+ try:
+ proc = await asyncio.create_subprocess_exec(
+ *cmd,
+ stdout=asyncio.subprocess.PIPE,
+ stderr=asyncio.subprocess.DEVNULL,
+ env={**os.environ, "XDG_RUNTIME_DIR": os.environ.get(
+ "XDG_RUNTIME_DIR", f"/run/user/{os.getuid()}")},
+ )
+ out, _ = await asyncio.wait_for(proc.communicate(), timeout=timeout)
+ if proc.returncode != 0:
+ return None
+ return out.decode("utf-8", "replace")
+ except (asyncio.TimeoutError, OSError):
+ return None
+
+
+def _detect_backend(base_url: str) -> str:
+ if ":11434" in base_url:
+ return "ollama"
+ if ":8001" in base_url:
+ return "llamacpp"
+ return "unknown"
+
+
+async def _llamacpp_running(container: str = "va_llm") -> bool:
+ out = await _run(["docker", "ps", "--filter", f"name=^{container}$", "--format", "{{.Names}}"])
+ return bool(out and container in out)
+
+
+async def _ollama_loaded() -> tuple[bool, list[dict]]:
+ """(dienst_erreichbar, [geladene Modelle])."""
+ out = await _run(["ollama", "ps"])
+ if out is None:
+ return False, []
+ models: list[dict] = []
+ lines = [ln for ln in out.splitlines() if ln.strip()]
+ for ln in lines[1:]: # Kopfzeile ueberspringen
+ # Spalten sind durch 2+ Leerzeichen getrennt: NAME ID SIZE PROCESSOR CONTEXT UNTIL
+ import re
+ cols = re.split(r"\s{2,}", ln.strip())
+ if cols:
+ models.append({
+ "name": cols[0],
+ "size": cols[2] if len(cols) > 2 else "",
+ "processor": cols[3] if len(cols) > 3 else "",
+ })
+ return True, models
+
+
+async def _gpus() -> list[dict]:
+ out = await _run([
+ "nvidia-smi",
+ "--query-gpu=index,memory.used,memory.total",
+ "--format=csv,noheader,nounits",
+ ])
+ if out is None:
+ return []
+ gpus: list[dict] = []
+ for ln in out.splitlines():
+ parts = [p.strip() for p in ln.split(",")]
+ if len(parts) == 3 and parts[0].isdigit():
+ used, total = int(parts[1]), int(parts[2])
+ gpus.append({
+ "index": int(parts[0]),
+ "used_mib": used,
+ "total_mib": total,
+ "percent": round(used / total * 100) if total else 0,
+ })
+ return gpus
+
+
+async def _gateway_service_active() -> bool:
+ out = await _run(["systemctl", "--user", "is-active", "voice-assistant.service"], timeout=4.0)
+ return bool(out and out.strip() == "active")
+
+
+async def llm_status() -> dict:
+ """Aggregierter, read-only LLM-/System-Status fuer das Admin-Panel."""
+ base_url = settings.local_llm_base_url
+ backend = _detect_backend(base_url)
+ llamacpp, (ollama_reachable, ollama_models), gpus, gw = await asyncio.gather(
+ _llamacpp_running(),
+ _ollama_loaded(),
+ _gpus(),
+ _gateway_service_active(),
+ )
+ return {
+ "backend": backend,
+ "model": settings.local_llm_model,
+ "base_url": base_url,
+ "llamacpp_running": llamacpp,
+ "ollama_reachable": ollama_reachable,
+ "ollama_loaded": ollama_models,
+ "gpus": gpus,
+ "gateway_service_active": gw,
+ }
diff --git a/app/api/admin.py b/app/api/admin.py
index ece441d..a862d5a 100644
--- a/app/api/admin.py
+++ b/app/api/admin.py
@@ -6,6 +6,7 @@ from fastapi import APIRouter, Depends, HTTPException, Request, WebSocket, WebSo
from fastapi.responses import FileResponse
from pydantic import BaseModel
+from app.admin_llm import llm_status
from app.auth import is_admin_user, require_admin, require_admin_or_user
from app.config import settings
from app.dependencies import get_store
@@ -187,6 +188,14 @@ async def export_db():
)
+# ── LLM-/System-Status (read-only) ──────────────────────────────────────────
+
+@router.get("/admin/llm/status", dependencies=[Depends(require_admin)])
+async def get_llm_status():
+ """Read-only Status: aktives Backend, Modell, GPU-Auslastung, Dienste."""
+ return await llm_status()
+
+
# ── Aussprache-Lexikon CRUD ─────────────────────────────────────────────────
def _read_pronunciation(lang: str) -> dict:
diff --git a/app/web/app.js b/app/web/app.js
index f04bd9f..f330bf6 100644
--- a/app/web/app.js
+++ b/app/web/app.js
@@ -956,13 +956,56 @@ async function loadEmergencyEvents() {
// ════════════════════════════════════════
// TAB: STATUS
// ════════════════════════════════════════
+// LLM-/GPU-Status-Karte (read-only) für den Status-Tab.
+function renderLlmCard(llm) {
+ if (!llm) return "";
+ const backendLabel = { ollama: "Ollama", llamacpp: "llama.cpp", unknown: "—" }[llm.backend] || llm.backend;
+ const dot = (on) => ``;
+ const row = (label, val) =>
+ `${label}
+ ${escHtml(val ?? "—")}`;
+
+ const loaded = (llm.ollama_loaded || []).map((m) =>
+ `${escHtml(m.name)} (${escHtml(m.processor || "")})`
+ ).join(", ") || "—";
+
+ const gpuBars = (llm.gpus || []).map((g) => `
+
+
GPU ${g.index}
+
+
${g.used_mib} / ${g.total_mib} MiB
+
keine GPU-Daten (nvidia-smi)
';
+
+ return `
+
+
+ ${dot(true)}
LLM-Backend
+ ${escHtml(backendLabel)}
+
+
+ ${row("Modell", llm.model)}
+ ${row("URL", llm.base_url)}
+ ${row("Ollama erreichbar", llm.ollama_reachable ? "ja" : "nein")}
+ ${row("llama.cpp läuft", llm.llamacpp_running ? "ja" : "nein")}
+ ${row("Geladene Modelle", loaded)}
+ ${row("Gateway-Dienst", llm.gateway_service_active ? "aktiv (systemd)" : "Vordergrund/aus")}
+
+
+ ${gpuBars}
+
+
lade …
';
try {
- const [cfg, met] = await Promise.all([
+ const [cfg, met, llm] = await Promise.all([
fetch("/api/config").then((r) => r.json()),
fetch("/api/metrics").then((r) => r.json()),
+ adminFetch("/api/admin/llm/status"),
]);
const route = cfg.default_route || {};
const counters = met.counters || {};
@@ -1000,6 +1043,8 @@ async function loadStatus() {
+ ${renderLlmCard(llm)}
+
Laufzeit-Metriken
diff --git a/app/web/index.html b/app/web/index.html
index f17d0f3..f715f87 100644
--- a/app/web/index.html
+++ b/app/web/index.html
@@ -250,6 +250,6 @@
-
+