AI-Profile-Router/router/ai_profile_router.py

#!/usr/bin/env python3
"""AI Profile Router – OpenAI-kompatibler Proxy vor llama.cpp.

Leitet OpenAI-kompatible Requests transparent an den lokalen llama.cpp-Server
weiter (Streaming, Tool Calls, JSON) und schaltet zwischen drei festen
Profilen um:

    Profil   Kontext
    ------   --------
    fast     73728
    medium   94208
    long     131072

Virtuelle Modelle:  qwen-fast, qwen-medium, qwen-long
Kommandos:          POST /fast, /medium, /long   (Profilwechsel)
                    GET  /status                 (Zustand)

Bildgenerierung (FLUX.2 [klein] 4B Base):
                    POST /v1/images/generations  (OpenAI-kompatibel)
                    GET  /images                 (Liste)
                    GET  /images/<datei>         (PNG-Download)

Sprachausgabe (Kokoro-82M, deutsch, CPU-only):
                    POST /v1/audio/speech        (OpenAI-kompatibel)
                    GET  /v1/audio/voices        (verfügbare Stimmen)

Spracherkennung (whisper.cpp, deutsch, CPU-only):
                    POST /v1/audio/transcriptions (OpenAI-kompatibel)
                    GET  /v1/audio/models        (verfügbare Audio-Modelle)

Der TTS-Worker (mike-ai-kokoro.service) und der STT-Worker
(mike-ai-whisper.service) laufen als separate, langlebige Prozesse.
Der Router leitet /v1/audio/speech und /v1/audio/transcriptions
per HTTP an die Worker weiter.

Der Router agiert als Modell-Orchestrator: vor der Generierung wird
llama.cpp gestoppt, der Bild-Worker lädt FLUX, generiert und entlädt
das Modell wieder; danach wird das vorherige Qwen-Profil wiederher-
gestellt und erst dann geantwortet (try/finally – Qwen wird auch bei
Fehlgeschlagener Generierung wiederhergestellt).

Nur Python-Standardbibliothek. Logging nach stdout (journald).
"""

from __future__ import annotations

import base64
import email
import json
import logging
import os
import queue
import re
import subprocess
import sys
import threading
import time
import uuid
import http.client
from email.parser import BytesParser
from email.policy import compat32
from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer

# ---------------------------------------------------------------------------
# Konfiguration (über Umgebungsvariablen, vgl. systemd-Unit)
# ---------------------------------------------------------------------------

HOST = os.environ.get("ROUTER_HOST", "0.0.0.0")
PORT = int(os.environ.get("ROUTER_PORT", "8081"))
UPSTREAM_URL = os.environ.get("UPSTREAM_URL", "http://127.0.0.1:8080").rstrip("/")
PROFILE_SCRIPT = os.environ.get("PROFILE_SCRIPT", "/usr/local/bin/llama-profile")
PROFILE_DIR = os.environ.get(
    "PROFILE_DIR", "/etc/systemd/system/mike-ai-llama-ui.service.d")

SWITCH_TIMEOUT = float(os.environ.get("SWITCH_TIMEOUT", "600"))     # s, Warten auf llama.cpp
REQUEST_TIMEOUT = float(os.environ.get("REQUEST_TIMEOUT", "600"))   # s, Read-Timeout Upstream
CONNECT_TIMEOUT = float(os.environ.get("CONNECT_TIMEOUT", "10"))    # s, Connect-Timeout
POLL_INTERVAL = float(os.environ.get("POLL_INTERVAL", "2"))         # s, Polling-Intervall

# --- Bildgenerierung (FLUX.2 [klein] 4B Base) ---
LLAMA_SERVICE = os.environ.get("LLAMA_SERVICE", "mike-ai-llama-ui.service")
SYSTEMCTL_BIN = os.environ.get("SYSTEMCTL_BIN", "systemctl")
IMAGE_WORKER = os.environ.get(
    "IMAGE_WORKER", "/opt/mike-ai/ai-profile-router/image_worker.py")
IMAGE_PYTHON = os.environ.get(
    "IMAGE_PYTHON", "/opt/mike-ai/ai-profile-router/venv/bin/python")
IMAGE_DIR = os.environ.get(
    "IMAGE_DIR", "/opt/mike-ai/ai-profile-router/images")
IMAGE_WORKER_LOG = os.environ.get(
    "IMAGE_WORKER_LOG", "/opt/mike-ai/ai-profile-router/image_worker.log")
IMAGE_START_TIMEOUT = float(os.environ.get("IMAGE_START_TIMEOUT", "120"))   # s, Worker-Start
IMAGE_GEN_TIMEOUT = float(os.environ.get("IMAGE_GEN_TIMEOUT", "1800"))      # s, pro Bild
IMAGE_VRAM_FREE_TIMEOUT = float(os.environ.get("IMAGE_VRAM_FREE_TIMEOUT", "90"))  # s, VRAM-Abgabe

# Erlaubte Auflösungen (Breite x Höhe). FLUX.2 klein ist für 1 MP
# ausgelegt; 1920x1088 (≈2 MP) wird zusätzlich unterstützt.
IMAGE_SIZES = {
    "1024x1024": (1024, 1024),
    "1536x1024": (1536, 1024),
    "1024x1536": (1024, 1536),
    "1920x1088": (1920, 1088),
    "1088x1920": (1088, 1920),
}
# Qualitätsstufen → Inference-Schritte (guidance bleibt offiziell 4.0).
# Auf der RTX 5080 gemessen: 30 vs. 50 Steps liefern praktisch dieselbe
# Qualität (1024x1024: 31,3 s vs. 45,3 s). Default ist daher "standard".
IMAGE_QUALITY = {"standard": 30, "high": 50}
IMAGE_DEFAULT_QUALITY = "standard"
IMAGE_MAX_N = 4

# --- Sprachausgabe (Kokoro-82M, deutsch, CPU-only) ---
TTS_WORKER_URL = os.environ.get("TTS_WORKER_URL", "http://127.0.0.1:8082")
TTS_TIMEOUT = float(os.environ.get("TTS_TIMEOUT", "300"))  # s, pro Synthese
TTS_CONNECT_TIMEOUT = float(os.environ.get("TTS_CONNECT_TIMEOUT", "5"))
TTS_MODEL = "kokoro-german"  # virtuelles Modell für /v1/audio/speech
TTS_VOICES = ("martin", "victoria")
TTS_DEFAULT_VOICE = "martin"
TTS_FORMATS = ("mp3", "wav", "flac", "pcm")
TTS_DEFAULT_FORMAT = "mp3"

# --- Spracherkennung (whisper.cpp, deutsch, CPU-only) ---
STT_WORKER_URL = os.environ.get("STT_WORKER_URL", "http://127.0.0.1:8084")
STT_TIMEOUT = float(os.environ.get("STT_TIMEOUT", "120"))  # s, pro Transkription
STT_CONNECT_TIMEOUT = float(os.environ.get("STT_CONNECT_TIMEOUT", "5"))
STT_MODEL = "whisper-1"  # virtuelles Modell für /v1/audio/transcriptions

# Chat-Waiting: Während eines Image-Jobs oder Profilwechsels ist Qwen
# down. Chat-Requests warten (statt 502) bis Qwen wieder bereit ist.
CHAT_WAIT_TIMEOUT = float(os.environ.get("CHAT_WAIT_TIMEOUT", "300"))  # s, max. Warten
CHAT_DRAIN_TIMEOUT = float(os.environ.get("CHAT_DRAIN_TIMEOUT", "60"))  # s, max. Warten auf aktive Chats

PROFILES = {"fast": 73728, "medium": 94208, "long": 131072}
VIRTUAL_MODELS = {f"qwen-{name}": name for name in PROFILES}

log = logging.getLogger("ai-profile-router")

# Hop-by-hop-Header, die nicht an Upstream/Client weitergereicht werden.
HOP_BY_HOP = {
    "host", "connection", "keep-alive", "proxy-authenticate",
    "proxy-authorization", "te", "trailer", "transfer-encoding",
    "upgrade", "content-length",
}


def _parse_upstream(url: str) -> tuple[str, int]:
    """'http://127.0.0.1:8080' -> ('127.0.0.1', 8080)"""
    hostport = url.split("://", 1)[-1]
    host, _, port = hostport.partition(":")
    return host, int(port) if port else 80


UPSTREAM_HOST, UPSTREAM_PORT = _parse_upstream(UPSTREAM_URL)


# ---------------------------------------------------------------------------
# Zustand
# ---------------------------------------------------------------------------

class _ImageState:
    """Zustand der Bildgenerierung (nur für Status-Reporting)."""
    def __init__(self) -> None:
        self.phase = "idle"               # siehe PHASES unten
        self.worker: "_Worker | None" = None
        self.last_error: str | None = None
        self.last_image: str | None = None
        self.last_seconds: float | None = None


IMAGE_PHASES = (
    "idle", "stopping-qwen", "loading-image", "generating",
    "unloading-image", "restoring-qwen",
)


class _State:
    """Gemeinsamer, thread-sicherer Zustand.

    lock        : zentraler GPU-/Model-Lock. Wird von Profilwechsel UND
                  Image-Generation gehalten → gegenseitiger Ausschluss,
                  kein Race zwischen beiden.
    avail_lock  : schützt qwen_unavailable + active_chats (Chat-Waiting).
    """
    lock = threading.Lock()          # GPU-/Model-Lock (Profilwechsel + Image)
    switching: str | None = None     # Profil, das gerade gewechselt wird
    started = time.time()
    image = _ImageState()
    # Qwen-Verfügbarkeit für das Chat-Waiting:
    qwen_unavailable = False         # True, wenn Qwen down/neu geladen wird
    active_chats = 0                 # Anzahl laufender Chat-Requests
    avail_lock = threading.Lock()    # schützt die beiden Felder oben


STATE = _State()


def _wait_chats_drained(timeout: float | None = None) -> None:
    """Wartet, bis keine aktiven Chat-Requests mehr laufen.

    Wird von Profilwechsel/Image-Job aufgerufen, BEVOR Qwen gestoppt wird.
    Verhindert, dass ein laufender Chat auf ein gestopptes Qwen trifft (502).
    """
    timeout = CHAT_DRAIN_TIMEOUT if timeout is None else timeout
    deadline = time.monotonic() + timeout
    while True:
        with STATE.avail_lock:
            if STATE.active_chats == 0:
                return
            n = STATE.active_chats
        if time.monotonic() > deadline:
            log.warning("Chat-Drain-Timeout nach %.0f s (%d aktive Chats) – "
                        "fahre trotzdem fort", timeout, n)
            return
        time.sleep(0.5)


def _set_qwen_unavailable(unavailable: bool) -> None:
    with STATE.avail_lock:
        STATE.qwen_unavailable = unavailable


# ---------------------------------------------------------------------------
# Upstream (llama.cpp)
# ---------------------------------------------------------------------------

def tts_status() -> dict:
    """Prüft den TTS-Worker: erreichbar? bereit? welche Stimmen?"""
    hostport = TTS_WORKER_URL.split("://", 1)[-1]
    host, _, port = hostport.partition(":")
    try:
        conn = http.client.HTTPConnection(host, int(port) if port else 80,
                                          timeout=TTS_CONNECT_TIMEOUT)
        conn.request("GET", "/status")
        resp = conn.getresponse()
        data = json.loads(resp.read())
        conn.close()
        return {"reachable": True, **data}
    except (OSError, ValueError) as e:
        return {"reachable": False, "error": str(e)}


def tts_synthesize(text: str, voice: str, speed: float,
                   fmt: str) -> tuple[bytes, str]:
    """Synthetisiert Audio über den TTS-Worker.

    Liefert (audio_bytes, content_type). Wirft RuntimeError bei Fehler.
    """
    hostport = TTS_WORKER_URL.split("://", 1)[-1]
    host, _, port = hostport.partition(":")
    payload = json.dumps({"text": text, "voice": voice,
                          "speed": speed, "format": fmt}).encode()
    try:
        conn = http.client.HTTPConnection(host, int(port) if port else 80,
                                          timeout=TTS_CONNECT_TIMEOUT)
        conn.request("POST", "/tts", body=payload,
                     headers={"Content-Type": "application/json"})
        conn.sock.settimeout(TTS_TIMEOUT)
        resp = conn.getresponse()
        body = resp.read()
        conn.close()
    except (OSError, http.client.HTTPException) as e:
        raise RuntimeError(f"TTS-Worker nicht erreichbar: {e}")
    if resp.status != 200:
        try:
            err = json.loads(body)
            msg = err.get("error", str(err))
        except ValueError:
            msg = body.decode(errors="replace")[:200]
        raise RuntimeError(f"TTS-Fehler ({resp.status}): {msg}")
    content_type = {"mp3": "audio/mpeg", "wav": "audio/wav",
                    "flac": "audio/flac",
                    "pcm": "application/octet-stream"}[fmt]
    return body, content_type


def stt_status() -> dict:
    """Prüft den STT-Worker: erreichbar? bereit?"""
    hostport = STT_WORKER_URL.split("://", 1)[-1]
    host, _, port = hostport.partition(":")
    try:
        conn = http.client.HTTPConnection(host, int(port) if port else 80,
                                          timeout=STT_CONNECT_TIMEOUT)
        conn.request("GET", "/status")
        resp = conn.getresponse()
        data = json.loads(resp.read())
        conn.close()
        return {"reachable": True, **data}
    except (OSError, ValueError) as e:
        return {"reachable": False, "error": str(e)}


def stt_transcribe(file_data: bytes, filename: str,
                   language: str | None = None,
                   prompt: str | None = None,
                   temperature: float | None = None) -> dict:
    """Transkribiert Audio über den STT-Worker.

    Liefert dict mit 'text'. Wirft RuntimeError bei Fehler.
    """
    hostport = STT_WORKER_URL.split("://", 1)[-1]
    host, _, port = hostport.partition(":")

    # Multipart-Form-Data bauen
    boundary = "----STTBoundary" + uuid.uuid4().hex[:16]
    parts = []
    parts.append(
        f"--{boundary}\r\n"
        f'Content-Disposition: form-data; name="file"; filename="{filename}"\r\n'
        f"Content-Type: application/octet-stream\r\n\r\n".encode("utf-8")
    )
    parts.append(file_data)
    parts.append(b"\r\n")
    for key, value in [("language", language), ("prompt", prompt),
                       ("temperature", temperature)]:
        if value is not None:
            parts.append(
                f"--{boundary}\r\n"
                f'Content-Disposition: form-data; name="{key}"\r\n\r\n'
                f"{value}\r\n".encode("utf-8")
            )
    parts.append(f"--{boundary}--\r\n".encode("utf-8"))
    body = b"".join(parts)

    try:
        conn = http.client.HTTPConnection(host, int(port) if port else 80,
                                          timeout=STT_CONNECT_TIMEOUT)
        conn.request("POST", "/transcribe", body=body,
                     headers={"Content-Type":
                              f"multipart/form-data; boundary={boundary}"})
        conn.sock.settimeout(STT_TIMEOUT)
        resp = conn.getresponse()
        data = json.loads(resp.read())
        conn.close()
    except (OSError, http.client.HTTPException) as e:
        raise RuntimeError(f"STT-Worker nicht erreichbar: {e}")
    if resp.status != 200:
        msg = data.get("error", str(data)) if isinstance(data, dict) else str(data)
        raise RuntimeError(f"STT-Fehler ({resp.status}): {msg}")
    return data


def upstream_status() -> dict:
    """Prüft llama.cpp: erreichbar? welches Modell? welcher Kontext?"""
    try:
        conn = http.client.HTTPConnection(UPSTREAM_HOST, UPSTREAM_PORT,
                                          timeout=CONNECT_TIMEOUT)
        conn.request("GET", "/v1/models")
        resp = conn.getresponse()
        data = json.loads(resp.read())
        conn.close()
    except (OSError, ValueError) as e:
        return {"reachable": False, "error": str(e)}
    models = data.get("data") or []
    if not models:
        return {"reachable": True, "model": None, "ctx": None}
    m = models[0]
    return {"reachable": True,
            "model": m.get("id"),
            "ctx": (m.get("meta") or {}).get("n_ctx")}


# ---------------------------------------------------------------------------
# Profile
# ---------------------------------------------------------------------------

def _read(path: str) -> str:
    with open(path, encoding="utf-8") as f:
        return f.read().strip()


def current_profile() -> str | None:
    """Aktives Profil, ermittelt durch Vergleich der override.conf."""
    try:
        override = _read(os.path.join(PROFILE_DIR, "override.conf"))
    except OSError:
        return None
    for name in PROFILES:
        try:
            ref = _read(os.path.join(PROFILE_DIR, f"profile-{name}.conf.disabled"))
        except OSError:
            continue
        if override == ref:
            return name
    return None


def _wait_ready(profile: str, deadline: float) -> None:
    """Wartet, bis llama.cpp das Profil geladen hat (Modell + ctx)."""
    expected_ctx = PROFILES[profile]
    while True:
        status = upstream_status()
        if (status["reachable"] and status.get("model")
                and status.get("ctx") == expected_ctx):
            log.info("llama.cpp bereit: Profil=%s Modell=%s ctx=%s",
                     profile, status.get("model"), status.get("ctx"))
            return
        if time.monotonic() > deadline:
            raise RuntimeError(
                f"llama.cpp nach {SWITCH_TIMEOUT:.0f} s nicht bereit "
                f"(erwartet ctx {expected_ctx}, aktuell: {status.get('ctx')})")
        time.sleep(POLL_INTERVAL)


def switch_profile(profile: str, implicit: bool = False) -> None:
    """Stellt sicher, dass das Profil aktiv ist, und wartet bis es geladen ist.

    Wirft RuntimeError, wenn das Profil nicht aktiviert werden konnte.

    implicit=True (ausgelöst durch ein virtuelles Modell in einem Chat-Request):
    Wenn das Profil bereits aktiv ist, aber llama.cpp down ist, wird sofort
    eine RuntimeError geworfen (kein stiller Neustart). Der Nutzer kann den
    Neustart explizit über /<profil> anstoßen.
    """
    if profile not in PROFILES:
        raise ValueError(f"unbekanntes Profil: {profile!r} "
                         f"(erlaubt: {', '.join(PROFILES)})")
    # Kein Fast-Fail: Wenn ein Image-Job läuft (hält den GPU-Lock), wartet
    # der Profilwechsel auf den GPU-Lock (blockiert), bis der Image-Job
    # fertig ist. So bekommen Chat-Requests kein 502, sondern warten.
    with STATE.lock:
        STATE.switching = profile
        try:
            cur = current_profile()
            up = upstream_status()
            ready = (up["reachable"] and up.get("model")
                     and up.get("ctx") == PROFILES[profile])
            if cur == profile and ready:
                log.info("Profil %s ist bereits aktiv", profile)
                return
            # Qwen wird neu geladen/gewechselt → für Chats nicht verfügbar.
            _set_qwen_unavailable(True)
            try:
                _wait_chats_drained()
                if cur == profile and up["reachable"] and not ready:
                    # Modell wird gerade geladen (z.B. nach einem Wechsel)
                    log.info("Warte, bis Profil %s geladen ist ...", profile)
                    _wait_ready(profile, time.monotonic() + SWITCH_TIMEOUT)
                    return
                if cur == profile and not up["reachable"] and implicit:
                    raise RuntimeError(
                        f"llama.cpp nicht erreichbar (Profil {profile} ist "
                        f"bereits aktiv; Neustart über /{profile})")
                log.info("Profilwechsel: %s -> %s", cur, profile)
                try:
                    proc = subprocess.run(
                        [PROFILE_SCRIPT, profile],
                        stdin=subprocess.DEVNULL,
                        stdout=subprocess.PIPE,
                        stderr=subprocess.STDOUT,
                        timeout=120,
                    )
                    out = proc.stdout.decode(errors="replace").strip()
                    if out:
                        log.info("llama-profile: %s", out[-500:])
                    if proc.returncode != 0:
                        # whiptail bricht das Skript ohne TTY ab – der Wechsel
                        # selbst (cp + systemctl restart) ist dann erledigt.
                        log.warning("llama-profile Exit-Code %d (ohne TTY "
                                    "erwartet)", proc.returncode)
                except subprocess.TimeoutExpired:
                    log.error("llama-profile hat 120 s überschritten")
                if current_profile() != profile:
                    raise RuntimeError(
                        f"Profildatei wurde nicht gesetzt (erwartet: {profile})")
                log.info("Warte, bis llama.cpp das Profil geladen hat ...")
                _wait_ready(profile, time.monotonic() + SWITCH_TIMEOUT)
            finally:
                _set_qwen_unavailable(False)
        finally:
            STATE.switching = None


# ---------------------------------------------------------------------------
# Bildgenerierung (FLUX.2 [klein] 4B Base)
# ---------------------------------------------------------------------------

class _Worker:
    """Verwaltet den Bild-Worker-Prozess (stdin/stdout-JSON-Protokoll)."""

    def __init__(self) -> None:
        self.proc: subprocess.Popen | None = None
        self.model_loaded = False
        self._queue: queue.Queue[dict] = queue.Queue()
        self._reader: threading.Thread | None = None

    def alive(self) -> bool:
        return self.proc is not None and self.proc.poll() is None

    def start(self) -> None:
        if self.alive():
            return
        log.info("starte Bild-Worker: %s %s", IMAGE_PYTHON, IMAGE_WORKER)
        logf = open(IMAGE_WORKER_LOG, "ab")
        self.proc = subprocess.Popen(
            [IMAGE_PYTHON, IMAGE_WORKER],
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            stderr=logf,
            text=True,
            bufsize=1,
        )
        self._reader = threading.Thread(target=self._read_loop, daemon=True)
        self._reader.start()
        try:
            msg = self._queue.get(timeout=IMAGE_START_TIMEOUT)
        except queue.Empty:
            self.stop()
            raise RuntimeError("Bild-Worker hat nicht gestartet")
        if msg.get("status") != "ready":
            self.stop()
            raise RuntimeError(f"Bild-Worker-Startfehler: {msg}")
        log.info("Bild-Worker bereit")

    def _read_loop(self) -> None:
        assert self.proc is not None and self.proc.stdout is not None
        for line in self.proc.stdout:
            line = line.strip()
            if not line:
                continue
            try:
                self._queue.put(json.loads(line))
            except ValueError:
                log.warning("Worker-Zeile (kein JSON): %s", line[:200])

    def request(self, payload: dict, timeout: float) -> dict:
        if not self.alive():
            raise RuntimeError("Bild-Worker ist nicht aktiv")
        assert self.proc is not None and self.proc.stdin is not None
        self.proc.stdin.write(json.dumps(payload) + "\n")
        self.proc.stdin.flush()
        try:
            return self._queue.get(timeout=timeout)
        except queue.Empty:
            raise RuntimeError(
                f"Bild-Worker hat nach {timeout:.0f} s nicht geantwortet "
                f"(cmd={payload.get('cmd')})")

    def stop(self) -> None:
        if self.proc is not None and self.proc.poll() is None:
            self.proc.terminate()
            try:
                self.proc.wait(timeout=10)
            except subprocess.TimeoutExpired:
                self.proc.kill()
        self.proc = None
        self.model_loaded = False


def _worker() -> _Worker:
    """Worker-Instanz liefern (startet bei Bedarf)."""
    img = STATE.image
    if not img.worker or not img.worker.alive():
        if img.worker:
            img.worker.stop()
        img.worker = _Worker()
        img.worker.start()
    return img.worker


def _wait_upstream_down(deadline: float) -> None:
    """Wartet, bis llama.cpp den Port freigegeben hat (VRAM frei)."""
    while time.monotonic() < deadline:
        if not upstream_status()["reachable"]:
            return
        time.sleep(1)
    raise RuntimeError("llama.cpp gibt Port/VRAM nicht frei")


def _vram_used_mib() -> int | None:
    """Aktuelle VRAM-Belegung in MiB (via nvidia-smi), None bei Fehler."""
    try:
        out = subprocess.run(
            ["nvidia-smi", "--query-gpu=memory.used",
             "--format=csv,noheader,nounits"],
            stdin=subprocess.DEVNULL, stdout=subprocess.PIPE,
            stderr=subprocess.DEVNULL, timeout=10,
        ).stdout.decode().strip()
        return int(out.splitlines()[0].split()[0])
    except (OSError, ValueError, IndexError):
        return None


def _wait_vram_free(threshold_mib: int = 1000,
                    timeout: float | None = None) -> None:
    """Wartet, bis der VRAM unter threshold_mib fällt (FLUX entladen).

    Wird nach dem Beenden des Bild-Workers aufgerufen, um sicherzustellen,
    dass der VRAM (inkl. CUDA-Kontext) frei ist, bevor Qwen neu startet.
    Wenn nvidia-smi nicht verfügbar ist (z.B. lokale Tests), wird der
    Check übersprungen.
    """
    timeout = IMAGE_VRAM_FREE_TIMEOUT if timeout is None else timeout
    deadline = time.monotonic() + timeout
    last = _vram_used_mib()
    if last is None:
        log.info("VRAM-Check übersprungen (nvidia-smi nicht verfügbar)")
        return
    while time.monotonic() < deadline:
        if last <= threshold_mib:
            log.info("VRAM frei: %d MiB", last)
            return
        time.sleep(1)
        last = _vram_used_mib()
        if last is None:
            log.info("VRAM-Check übersprungen (nvidia-smi nicht verfügbar)")
            return
    raise RuntimeError(
        f"VRAM nach {timeout:.0f} s nicht frei (letzte Messung: "
        f"{last} MiB, erwartet <= {threshold_mib} MiB)")


def _restore_qwen(profile: str) -> None:
    """Startet llama.cpp mit dem gemerkten Profil und wartet auf Readiness."""
    log.info("stelle Qwen-Profil %s wieder her ...", profile)
    try:
        subprocess.run([SYSTEMCTL_BIN, "start", LLAMA_SERVICE],
                       stdin=subprocess.DEVNULL,
                       stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
                       timeout=120)
    except subprocess.TimeoutExpired:
        log.error("systemctl start hat 120 s überschritten")
    _wait_ready(profile, time.monotonic() + SWITCH_TIMEOUT)


def generate_image(prompt: str, width: int, height: int, steps: int,
                   guidance: float, seed: int | None, n: int,
                   quality: str = "standard"
                   ) -> tuple[list[str], str | None]:
    """Orchestriert die Bildgenerierung inkl. Qwen-Hotswap.

    Hält den zentralen GPU-Lock (gegenseitiger Ausschluss mit Profilwechsel).
    Ablauf: Qwen stoppen → Worker laden → generieren → Worker beenden
    (VRAM + CUDA-Kontext frei) → Qwen wiederherstellen. Qwen wird auch bei
    Fehlern wiederhergestellt (try/finally).
    """
    img = STATE.image
    with STATE.lock:
        if img.phase != "idle":
            raise RuntimeError(f"Bildgenerierung läuft ({img.phase})")
        profile = current_profile()
        if profile is None:
            raise RuntimeError("kein aktives Qwen-Profil (override.conf?)")
        os.makedirs(IMAGE_DIR, exist_ok=True)
        results: list[str] = []
        warning: str | None = None
        # Qwen wird gestoppt → für Chats nicht verfügbar (die warten).
        _set_qwen_unavailable(True)
        try:
            _wait_chats_drained()

            # 1) Qwen stoppen (VRAM freigeben).
            img.phase = "stopping-qwen"
            subprocess.run([SYSTEMCTL_BIN, "stop", LLAMA_SERVICE],
                           stdin=subprocess.DEVNULL,
                           stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
                           timeout=120)
            _wait_upstream_down(time.monotonic() + 60)

            # 2) Worker starten (Modell wird beim ersten generate geladen).
            img.phase = "loading-image"
            worker = _worker()

            # 3) Generieren.
            for i in range(n):
                img.phase = "generating"
                filename = time.strftime("%Y%m%d-%H%M%S") + \
                    f"-{os.urandom(2).hex()}.png"
                output = os.path.join(IMAGE_DIR, filename)
                resp = worker.request({
                    "cmd": "generate",
                    "prompt": prompt,
                    "width": width,
                    "height": height,
                    "steps": steps,
                    "guidance": guidance,
                    "seed": seed,
                    "output": output,
                }, timeout=IMAGE_GEN_TIMEOUT)
                if resp.get("status") != "ok":
                    raise RuntimeError(
                        resp.get("message", "Bildgenerierung fehlgeschlagen"))
                worker.model_loaded = True
                results.append(filename)
                img.last_image = filename
                img.last_seconds = resp.get("seconds")
                # Metadaten speichern (Sidecar-JSON).
                meta = {
                    "prompt": prompt,
                    "seed": seed,
                    "width": width,
                    "height": height,
                    "size": f"{width}x{height}",
                    "steps": steps,
                    "guidance": guidance,
                    "quality": quality,
                    "seconds": resp.get("seconds"),
                    "model": "FLUX.2-klein-base-4B",
                    "created": time.strftime("%Y-%m-%dT%H:%M:%S"),
                }
                meta_path = os.path.join(IMAGE_DIR, filename[:-4] + ".json")
                try:
                    with open(meta_path, "w", encoding="utf-8") as f:
                        json.dump(meta, f, ensure_ascii=False, indent=2)
                except OSError as e:
                    log.warning("Metadaten-Speicherung fehlgeschlagen: %s", e)
                log.info("Bild %d/%d: %s (%.1f s)", i + 1, n, filename,
                         resp.get("seconds", 0))

            # 4) Worker vollständig beenden (VRAM + CUDA-Kontext freigeben).
            img.phase = "unloading-image"
            worker.stop()
            img.worker = None
            try:
                _wait_vram_free()
            except RuntimeError as e:
                log.warning("VRAM-Check: %s (fahre mit Qwen-Restore fort)", e)
        except Exception as e:
            img.last_error = str(e)
            log.error("Bildgenerierung fehlgeschlagen: %s", e)
            # Worker sicher beenden (falls noch aktiv), VRAM freigeben.
            if img.worker is not None:
                img.worker.stop()
                img.worker = None
            raise
        finally:
            # 5) Qwen immer wiederherstellen.
            img.phase = "restoring-qwen"
            try:
                _restore_qwen(profile)
                _set_qwen_unavailable(False)
            except Exception as e:
                warning = f"Qwen-Wiederherstellung fehlgeschlagen: {e}"
                img.last_error = warning
                log.error(warning)
                # Qwen ist down → qwen_unavailable bleibt True.
            img.phase = "idle"
    return results, warning


def _image_filename_ok(name: str) -> bool:
    return bool(re.fullmatch(r"[A-Za-z0-9][A-Za-z0-9._-]*\.png", name))


# ---------------------------------------------------------------------------
# HTTP-Handler
# ---------------------------------------------------------------------------

class Handler(BaseHTTPRequestHandler):
    server_version = "AIProfileRouter/1.0"
    timeout = 60  # Socket-Timeout für Client-Requests (s)

    # ---------- Routing ----------

    def do_GET(self):
        self._route()

    def do_POST(self):
        self._route()

    def _route(self):
        path = self.path.split("?", 1)[0]
        started = time.monotonic()
        try:
            if path == "/v1/models" and self.command == "GET":
                self._send_json(200, self._models_payload())
            elif path == "/status":
                self._send_json(200, self._status_payload())
            elif path == "/v1/audio/models" and self.command == "GET":
                self._send_json(200, self._audio_models_payload())
            elif path == "/v1/audio/voices" and self.command == "GET":
                self._send_json(200, self._audio_voices_payload())
            elif path == "/v1/images/generations" and self.command == "POST":
                self._image_generate()
            elif path == "/v1/audio/speech" and self.command == "POST":
                self._speech()
            elif path == "/v1/audio/transcriptions" and self.command == "POST":
                self._transcribe()
            elif path == "/images" and self.command == "GET":
                self._images_list()
            elif path.startswith("/images/") and self.command == "GET":
                self._image_serve(path[len("/images/"):])
            elif path in ("/fast", "/medium", "/long"):
                self._switch(path[1:])
            elif (self.command == "POST" and path.startswith("/")
                    and path.count("/") == 1):
                # Kommandonamensraum: unbekanntes Profil
                self._send_error(400, f"unbekanntes Profil: {path[1:]}",
                                 "invalid_request_error", "invalid_profile")
            else:
                self._forward()
        except BrokenPipeError:
            log.warning("Client getrennt: %s %s", self.command, path)
        except Exception:
            log.exception("Fehler bei %s %s", self.command, path)
            self._safe_error(500, "interner Router-Fehler")
        finally:
            log.info("%s %s -> %s in %.3f s", self.command, path,
                     getattr(self, "_last_code", "-"), time.monotonic() - started)

    # ---------- Router-eigene Endpunkte ----------

    @staticmethod
    def _models_payload() -> dict:
        return {
            "object": "list",
            "data": [
                {
                    "id": f"qwen-{name}",
                    "object": "model",
                    "created": 0,
                    "owned_by": "ai-profile-router",
                    "context_length": ctx,
                    "context_window": ctx,
                }
                for name, ctx in PROFILES.items()
            ],
        }

    def _status_payload(self) -> dict:
        up = upstream_status()
        img = STATE.image
        with STATE.avail_lock:
            qwen_unavailable = STATE.qwen_unavailable
            active_chats = STATE.active_chats
        return {
            "router": "ai-profile-router",
            "uptime_seconds": round(time.time() - STATE.started, 1),
            "current_profile": current_profile(),
            "switching": STATE.switching,
            "profiles": PROFILES,
            "upstream": {
                "url": UPSTREAM_URL,
                "reachable": up["reachable"],
                "model": up.get("model"),
                "ctx": up.get("ctx"),
            },
            "qwen": {
                "available": not qwen_unavailable,
                "active_chats": active_chats,
            },
            "image": {
                "phase": img.phase,
                "worker": "running" if (img.worker and img.worker.alive())
                          else "stopped",
                "model_loaded": bool(img.worker and img.worker.model_loaded),
                "last_image": img.last_image,
                "last_seconds": img.last_seconds,
                "last_error": img.last_error,
            },
            "tts": tts_status(),
            "stt": stt_status(),
        }

    # ---------- Bildgenerierung ----------

    def _image_generate(self) -> None:
        length = int(self.headers.get("Content-Length") or 0)
        try:
            data = json.loads(self.rfile.read(length))
        except ValueError:
            self._send_error(400, "ungültiges JSON",
                             "invalid_request_error", "invalid_json")
            return
        if not isinstance(data, dict):
            self._send_error(400, "Request muss ein JSON-Objekt sein",
                             "invalid_request_error", "invalid_request")
            return

        prompt = data.get("prompt")
        if not isinstance(prompt, str) or not prompt.strip():
            self._send_error(400, "'prompt' fehlt oder ist leer",
                             "invalid_request_error", "missing_prompt")
            return
        if len(prompt) > 8000:
            self._send_error(400, "'prompt' zu lang (max 8000 Zeichen)",
                             "invalid_request_error", "prompt_too_long")
            return

        # Größe
        size = data.get("size", "1024x1024")
        if size not in IMAGE_SIZES:
            self._send_error(
                400, f"ungültige Größe: {size!r} "
                     f"(erlaubt: {', '.join(IMAGE_SIZES)})",
                "invalid_request_error", "invalid_size")
            return
        width, height = IMAGE_SIZES[size]

        # Anzahl
        n = data.get("n", 1)
        if not isinstance(n, int) or isinstance(n, bool) or not 1 <= n <= IMAGE_MAX_N:
            self._send_error(400, f"'n' muss eine Ganzzahl 1..{IMAGE_MAX_N} sein",
                             "invalid_request_error", "invalid_n")
            return

        # Qualität / Schritte / Guidance
        quality = data.get("quality", IMAGE_DEFAULT_QUALITY)
        if quality not in IMAGE_QUALITY:
            self._send_error(400, f"ungültige Qualität: {quality!r} "
                                  f"(erlaubt: {', '.join(IMAGE_QUALITY)})",
                             "invalid_request_error", "invalid_quality")
            return
        steps = data.get("steps", IMAGE_QUALITY[quality])
        if not isinstance(steps, int) or isinstance(steps, bool) or not 4 <= steps <= 150:
            self._send_error(400, "'steps' muss eine Ganzzahl 4..150 sein",
                             "invalid_request_error", "invalid_steps")
            return
        guidance = data.get("guidance", 4.0)
        try:
            guidance = float(guidance)
        except (TypeError, ValueError):
            self._send_error(400, "'guidance' muss eine Zahl sein",
                             "invalid_request_error", "invalid_guidance")
            return
        if not 1.0 <= guidance <= 10.0:
            self._send_error(400, "'guidance' muss zwischen 1.0 und 10.0 sein",
                             "invalid_request_error", "invalid_guidance")
            return

        seed = data.get("seed")
        if seed is not None:
            try:
                seed = int(seed)
            except (TypeError, ValueError):
                self._send_error(400, "'seed' muss eine Ganzzahl sein",
                                 "invalid_request_error", "invalid_seed")
                return
            if not 0 <= seed <= 2**32 - 1:
                self._send_error(400, "'seed' muss zwischen 0 und 4294967295 sein",
                                 "invalid_request_error", "invalid_seed")
                return

        response_format = data.get("response_format", "url")
        if response_format not in ("url", "b64_json"):
            self._send_error(400, "'response_format' muss 'url' oder 'b64_json' sein",
                             "invalid_request_error", "invalid_response_format")
            return

        # Generierung (blockt mehrere Minuten – eigener Thread-Timeout).
        self.timeout = None
        try:
            results, warning = generate_image(
                prompt.strip(), width, height, steps, guidance, seed, n,
                quality)
        except (ValueError, RuntimeError) as e:
            self._send_error(503, str(e), "server_error", "image_generation_failed")
            return

        # Antwort bauen
        host = self.headers.get("Host") or f"{HOST}:{PORT}"
        if not host.startswith(("http://", "https://")):
            host = f"http://{host}"
        items = []
        for filename in results:
            path = os.path.join(IMAGE_DIR, filename)
            item: dict = {"url": f"{host}/images/{filename}", "b64_json": None}
            if response_format == "b64_json":
                with open(path, "rb") as f:
                    item["b64_json"] = base64.b64encode(f.read()).decode()
                item["url"] = None
            items.append(item)
        payload: dict = {"created": int(time.time()), "data": items}
        if warning:
            payload["router_warning"] = warning
        self._send_json(200, payload)

    def _images_list(self) -> None:
        if not os.path.isdir(IMAGE_DIR):
            self._send_json(200, {"images": []})
            return
        entries = []
        for name in sorted(os.listdir(IMAGE_DIR), reverse=True):
            if not _image_filename_ok(name):
                continue
            path = os.path.join(IMAGE_DIR, name)
            try:
                st = os.stat(path)
            except OSError:
                continue
            entry = {
                "name": name,
                "url": f"/images/{name}",
                "bytes": st.st_size,
                "modified": int(st.st_mtime),
            }
            # Metadaten laden (Sidecar-JSON, falls vorhanden).
            meta_path = os.path.join(IMAGE_DIR, name[:-4] + ".json")
            if os.path.isfile(meta_path):
                try:
                    with open(meta_path, encoding="utf-8") as f:
                        entry["meta"] = json.load(f)
                except (OSError, ValueError):
                    pass
            entries.append(entry)
        self._send_json(200, {"images": entries[:200]})

    def _image_serve(self, name: str) -> None:
        if not _image_filename_ok(name):
            self._send_error(400, "ungültiger Dateiname",
                             "invalid_request_error", "invalid_filename")
            return
        path = os.path.join(IMAGE_DIR, name)
        if not os.path.isfile(path):
            self._send_error(404, "Bild nicht gefunden",
                             "invalid_request_error", "not_found")
            return
        data = open(path, "rb").read()
        self._last_code = 200
        self.send_response(200)
        self.send_header("Content-Type", "image/png")
        self.send_header("Content-Length", str(len(data)))
        self.send_header("Cache-Control", "public, max-age=86400")
        self.send_header("Connection", "close")
        self.end_headers()
        self.wfile.write(data)

    # ---------- Sprachausgabe (Kokoro) ----------

    def _speech(self) -> None:
        length = int(self.headers.get("Content-Length") or 0)
        try:
            data = json.loads(self.rfile.read(length))
        except ValueError:
            self._send_error(400, "ungültiges JSON",
                             "invalid_request_error", "invalid_json")
            return
        if not isinstance(data, dict):
            self._send_error(400, "Request muss ein JSON-Objekt sein",
                             "invalid_request_error", "invalid_request")
            return

        # input (OpenAI) – auch 'text' akzeptieren (bequemer für curl)
        text = data.get("input", data.get("text"))
        if not isinstance(text, str) or not text.strip():
            self._send_error(400, "'input' fehlt oder ist leer",
                             "invalid_request_error", "missing_input")
            return
        if len(text) > 8000:
            self._send_error(400, "'input' zu lang (max 8000 Zeichen)",
                             "invalid_request_error", "input_too_long")
            return

        voice = data.get("voice", TTS_DEFAULT_VOICE)
        if voice not in TTS_VOICES:
            self._send_error(
                400, f"ungültige Stimme: {voice!r} "
                     f"(erlaubt: {', '.join(TTS_VOICES)})",
                "invalid_request_error", "invalid_voice")
            return

        fmt = data.get("response_format", TTS_DEFAULT_FORMAT)
        if fmt not in TTS_FORMATS:
            self._send_error(
                400, f"ungültiges response_format: {fmt!r} "
                     f"(erlaubt: {', '.join(TTS_FORMATS)})",
                "invalid_request_error", "invalid_format")
            return

        speed = data.get("speed", 1.0)
        try:
            speed = float(speed)
        except (TypeError, ValueError):
            self._send_error(400, "'speed' muss eine Zahl sein",
                             "invalid_request_error", "invalid_speed")
            return
        if not 0.5 <= speed <= 2.0:
            self._send_error(400, "'speed' muss zwischen 0.5 und 2.0 sein",
                             "invalid_request_error", "invalid_speed")
            return

        # Modell-Name optional; falls angegeben, muss es kokoro-german sein.
        model = data.get("model")
        if model is not None and model != TTS_MODEL:
            self._send_error(400, f"unbekanntes Modell: {model!r} "
                                  f"(erwartet: {TTS_MODEL})",
                             "invalid_request_error", "unknown_model")
            return

        self.timeout = None  # Synthese kann dauern
        try:
            audio, content_type = tts_synthesize(
                text.strip(), voice, speed, fmt)
        except RuntimeError as e:
            self._send_error(503, str(e), "server_error", "tts_failed")
            return

        self._last_code = 200
        self.send_response(200)
        self.send_header("Content-Type", content_type)
        self.send_header("Content-Length", str(len(audio)))
        self.send_header("Connection", "close")
        self.end_headers()
        self.wfile.write(audio)

    # ---------- Audio-Discovery ----------

    def _audio_models_payload(self) -> dict:
        """Listet verfügbare Audio-Modelle (STT + TTS)."""
        tts = tts_status()
        stt = stt_status()
        models = []
        if stt.get("ready"):
            models.append({
                "id": STT_MODEL,
                "object": "model",
                "owned_by": "whisper.cpp",
                "type": "transcription",
            })
        if tts.get("ready"):
            models.append({
                "id": TTS_MODEL,
                "object": "model",
                "owned_by": "kokoro",
                "type": "speech",
            })
        return {"object": "list", "data": models}

    def _audio_voices_payload(self) -> dict:
        """Listet verfügbare TTS-Stimmen."""
        tts = tts_status()
        voices = []
        for v in tts.get("voices", []):
            voices.append({
                "id": v,
                "object": "voice",
                "language": "de",
            })
        return {"object": "list", "data": voices}

    # ---------- STT (Spracherkennung) ----------

    def _parse_multipart(self, data: bytes, content_type: str
                         ) -> tuple[bytes, str, dict]:
        """Parst multipart/form-data. Liefert (file_data, filename, fields).

        Nutzt email.parser.BytesParser (Standardbibliothek) für robustes
        MIME-Parsing. Handhabt quoted und unquoted Boundaries, beliebige
        Feldreihenfolge, zusätzliche Header und binäre Payloads.
        """
        # MIME-Message aus rohen Bytes + Content-Type-Header bauen
        raw = (f"Content-Type: {content_type}\r\n\r\n"
               ).encode("utf-8") + data
        msg = BytesParser(policy=compat32).parsebytes(raw)
        if not msg.is_multipart():
            raise ValueError("Kein multipart/form-data")

        file_data = b""
        filename = ""
        fields = {}

        for part in msg.get_payload():
            disposition = part.get("Content-Disposition", "")
            name = None
            part_filename = None
            for kv in disposition.split(";"):
                kv = kv.strip()
                if kv.startswith("name="):
                    name = kv[len("name="):].strip('"')
                elif kv.startswith("filename="):
                    part_filename = kv[len("filename="):].strip('"')
            if name is None:
                continue

            payload = part.get_payload(decode=True)
            if payload is None:
                payload = b""

            if part_filename is not None:
                # Dateifeld (binär, nicht dekodieren)
                file_data = payload
                filename = part_filename or ""
            else:
                # Textfeld
                fields[name] = payload.decode("utf-8", errors="replace")

        return file_data, filename, fields

    def _transcribe(self) -> None:
        """POST /v1/audio/transcriptions – STT (OpenAI-kompatibel)."""
        content_type = self.headers.get("Content-Type", "")
        if "multipart/form-data" not in content_type:
            self._send_error(400,
                             "Content-Type muss multipart/form-data sein",
                             "invalid_request_error", "invalid_content_type")
            return

        length = int(self.headers.get("Content-Length") or 0)
        data = self.rfile.read(length)

        try:
            file_data, filename, fields = self._parse_multipart(
                data, content_type)
        except ValueError as e:
            self._send_error(400, str(e),
                             "invalid_request_error", "invalid_multipart")
            return

        if not file_data:
            self._send_error(400, "Keine Datei im Request",
                             "invalid_request_error", "missing_file")
            return

        # Modell-Validierung
        model = fields.get("model", STT_MODEL)
        if model not in (STT_MODEL, "whisper"):
            self._send_error(400, f"unbekanntes Modell: {model!r} "
                                  f"(erwartet: {STT_MODEL})",
                             "invalid_request_error", "unknown_model")
            return

        # Optionale Felder
        language = fields.get("language")
        prompt = fields.get("prompt")
        temperature = None
        if fields.get("temperature"):
            try:
                temperature = float(fields["temperature"])
            except ValueError:
                self._send_error(400, "'temperature' muss eine Zahl sein",
                                 "invalid_request_error", "invalid_temperature")
                return
        response_format = fields.get("response_format", "json")

        self.timeout = None  # Transkription kann dauern
        try:
            result = stt_transcribe(
                file_data, filename,
                language=language, prompt=prompt,
                temperature=temperature)
        except RuntimeError as e:
            self._send_error(503, str(e), "server_error", "stt_failed")
            return

        # OpenAI-kompatibles Antwort-Format
        if response_format == "verbose_json":
            resp = {
                "text": result.get("text", ""),
                "language": result.get("language", "de"),
                "duration": result.get("audio_duration_ms", 0) / 1000.0,
            }
        else:
            resp = {"text": result.get("text", "")}
        self._send_json(200, resp)

    def _switch(self, profile: str) -> None:
        if profile not in PROFILES:
            self._send_error(400, f"unbekanntes Profil: {profile}",
                             "invalid_request_error", "invalid_profile")
            return
        try:
            switch_profile(profile)
        except (ValueError, RuntimeError) as e:
            self._send_error(503, str(e), "server_error", "profile_switch_failed")
            return
        up = upstream_status()
        self._send_json(200, {
            "status": "ok",
            "profile": profile,
            "context_length": PROFILES[profile],
            "model": up.get("model"),
        })

    # ---------- Transparentes Forwarding ----------

    def _forward(self) -> None:
        length = int(self.headers.get("Content-Length") or 0)
        body = self.rfile.read(length) if length > 0 else None

        # Virtuelles Modell? -> Profil sicherstellen, dann Modell ersetzen.
        if body is not None and self.path.startswith("/v1/"):
            try:
                data = json.loads(body)
            except ValueError:
                data = None
            model = data.get("model") if isinstance(data, dict) else None
            if isinstance(model, str) and model in VIRTUAL_MODELS:
                profile = VIRTUAL_MODELS[model]
                try:
                    switch_profile(profile, implicit=True)
                except (ValueError, RuntimeError) as e:
                    self._send_error(502, str(e), "server_error",
                                     "upstream_unavailable")
                    return
                up = upstream_status()
                if not up["reachable"] or not up.get("model"):
                    self._send_error(502, "llama.cpp nicht erreichbar",
                                     "server_error", "upstream_unavailable")
                    return
                data["model"] = up["model"]
                body = json.dumps(data).encode()
            elif isinstance(model, str) and model.startswith("qwen-"):
                # qwen-* ist der Namensraum des Routers
                self._send_error(400, f"unbekanntes virtuelles Modell: {model}",
                                 "invalid_request_error", "unknown_model")
                return

        # An llama.cpp weiterleiten (mit Chat-Waiting, Streaming bleibt erhalten).
        self._proxy_with_wait(body)

    def _proxy_with_wait(self, body: bytes | None) -> None:
        """Leitet an llama.cpp weiter, wartet aber erst, bis Qwen verfügbar ist.

        Während eines Image-Jobs oder Profilwechsels ist Qwen down. Statt
        502 zu liefern, wartet der Request (mit Timeout), bis Qwen wieder
        bereit ist. Mehrere Chats können parallel laufen (active_chats).

        Race-frei: Der Check auf qwen_unavailable und das Inkrement von
        active_chats sind atomar (avail_lock). Ein Image-Job/Profilwechsel
        setzt qwen_unavailable=True und wartet auf active_chats==0, BEVOR
        er Qwen stoppt – ein laufender Chat wird daher nie unterbrochen.
        """
        deadline = time.monotonic() + CHAT_WAIT_TIMEOUT
        while True:
            with STATE.avail_lock:
                if not STATE.qwen_unavailable:
                    STATE.active_chats += 1
                    break
            if time.monotonic() > deadline:
                self._send_error(
                    503,
                    "Qwen wird neu geladen (Image-Job oder Profilwechsel), "
                    "bitte später erneut",
                    "server_error", "qwen_reloading")
                return
            time.sleep(0.5)
        try:
            self._proxy(body)
        finally:
            with STATE.avail_lock:
                STATE.active_chats -= 1

    def _proxy(self, body: bytes | None) -> None:
        # An llama.cpp weiterleiten (Streaming bleibt erhalten).
        try:
            conn = http.client.HTTPConnection(UPSTREAM_HOST, UPSTREAM_PORT,
                                              timeout=CONNECT_TIMEOUT)
            conn.connect()
            conn.sock.settimeout(REQUEST_TIMEOUT)
            headers = {k: v for k, v in self.headers.items()
                       if k.lower() not in HOP_BY_HOP}
            conn.request(self.command, self.path, body=body, headers=headers)
            resp = conn.getresponse()
        except (OSError, http.client.HTTPException) as e:
            self._send_error(502, f"llama.cpp nicht erreichbar: {e}",
                             "server_error", "upstream_unavailable")
            return

        self._last_code = resp.status
        self.send_response(resp.status)
        for k, v in resp.getheaders():
            if k.lower() not in HOP_BY_HOP:
                self.send_header(k, v)
        self.send_header("Connection", "close")
        self.end_headers()
        try:
            while True:
                chunk = resp.read(16384)
                if not chunk:
                    break
                self.wfile.write(chunk)
                self.wfile.flush()
        except (OSError, http.client.HTTPException) as e:
            log.warning("Upstream-Stream abgebrochen: %s", e)
        finally:
            conn.close()

    # ---------- Antworten ----------

    def _send_json(self, code: int, payload: dict) -> None:
        body = json.dumps(payload).encode()
        self._last_code = code
        self.send_response(code)
        self.send_header("Content-Type", "application/json")
        self.send_header("Content-Length", str(len(body)))
        self.send_header("Connection", "close")
        self.end_headers()
        self.wfile.write(body)

    def _send_error(self, code: int, message: str, etype: str, ecode: str) -> None:
        # OpenAI-kompatibles Fehlerformat
        self._send_json(code, {"error": {"message": message,
                                         "type": etype,
                                         "code": ecode}})

    def _safe_error(self, code: int, message: str) -> None:
        try:
            self._send_error(code, message, "server_error", "internal_error")
        except Exception:
            pass


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

class _FlushHandler(logging.StreamHandler):
    """StreamHandler, der nach jedem Record flusht (journald)."""

    def emit(self, record):
        super().emit(record)
        self.flush()


def main() -> None:
    handler = _FlushHandler(sys.stdout)
    handler.setFormatter(logging.Formatter(
        "%(asctime)s %(levelname)s %(message)s"))
    logging.basicConfig(level=os.environ.get("LOG_LEVEL", "INFO"),
                        handlers=[handler])
    log.info("AI Profile Router startet: %s:%s -> %s (Profile: %s)",
             HOST, PORT, UPSTREAM_URL, ", ".join(PROFILES))
    server = ThreadingHTTPServer((HOST, PORT), Handler)
    server.daemon_threads = True
    try:
        server.serve_forever()
    except KeyboardInterrupt:
        pass
    finally:
        server.server_close()


if __name__ == "__main__":
    main()