duffyduckandClaude Opus 4.8
5c25d6abeb
Neue lokale GGUF-Modelle per Knopf auf eine Box laden — ohne Image-Rebuild. - llm-adapter besitzt jetzt llama-swaps Config: generiert /models/llama-swap.config.yaml aus Basis-Template (xtts/llama-swap/config.yaml) + persistenter Registry /models/aria_models.json. Neue RVS-Handler llm_provision_model / llm_remove_model (targetInstance-gefiltert): Registry+ Config schreiben, llama-swap-Reload anstossen, neu announcen, Warmup (zieht das GGUF via -hf, Fortschritt via service_status loading→ready). pyyaml ergaenzt. - compose: llama-swap liest --config /models/llama-swap.config.yaml; llm-adapter mountet ./models (rw) + ./llama-swap (ro Template). - diagnostic/server.js: /shared/config/llm_catalog.json (kuratierte GGUF-Liste) + GET /api/llm-catalog + POST /api/llm-catalog/refresh (HuggingFace-API-Merge); Actions llm_provision_model / llm_remove_model / llm_test; llm_provision_result an Browser durchgereicht. - diagnostic/index.html: "Modell-Katalog"-Card (HF-Refresh, Ziel-Box waehlen, Laden, Verfuegbarkeit) + Test-Chat-Zeile ans lokale LLM (Antwort + Latenz). Download nutzt llama-swaps vorhandenen -hf-Pfad (kein neuer Download-Code). Reload ist der einzige Deploy-Verify-Punkt (llama-swap-Image); Fallback Box-up. Deploy: diagnostic neu bauen (VM) + llm-Boxen neu bauen. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
447 lines
18 KiB
Python
447 lines
18 KiB
Python
"""
|
|
ARIA Local-LLM-Adapter (AI-Box) — Plan B, Phase B0.
|
|
|
|
Bruecke zwischen RVS und dem lokalen llama.cpp-Server. Spiegelt das Muster der
|
|
whisper-bridge: verbindet sich per WebSocket mit dem RVS (Token-Room, TLS mit
|
|
ws-Fallback, Reconnect-Backoff), lauscht auf `llm_request` und ruft den lokalen
|
|
llama.cpp-`/v1/chat/completions`-Endpoint (OpenAI-kompatibel), antwortet mit
|
|
`llm_response` (korreliert per requestId).
|
|
|
|
Topologie: AI-Box steht zuhause, ARIA im RZ — die Kommunikation laeuft ueber
|
|
den RVS (wie TTS/STT), keine IPs zu pflegen. Nur URL + Token.
|
|
|
|
Env:
|
|
RVS_HOST, RVS_PORT, RVS_TLS, RVS_TLS_FALLBACK, RVS_TOKEN (wie f5tts/whisper)
|
|
LLAMA_URL Default http://llama:8081 (llama.cpp im selben Compose-Netz)
|
|
LLM_MODEL optionaler Modell-Name fuer llama (llama.cpp ignoriert ihn
|
|
meist, dient nur der Transparenz im Log)
|
|
LLM_TIMEOUT_SEC Default 60
|
|
|
|
Bewusst NICHT-streamend in B0 (volle llm_response). Token-Streaming (llm_partial)
|
|
kommt in B2 zusammen mit TTS-on-first-sentence.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import asyncio
|
|
import json
|
|
import logging
|
|
import os
|
|
import time
|
|
|
|
import httpx
|
|
import websockets
|
|
|
|
logging.basicConfig(
|
|
level=logging.INFO,
|
|
format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
|
|
)
|
|
logger = logging.getLogger("llm-adapter")
|
|
|
|
RVS_HOST = os.getenv("RVS_HOST", "").strip()
|
|
RVS_PORT = os.getenv("RVS_PORT", "443").strip()
|
|
RVS_TLS = os.getenv("RVS_TLS", "true").lower() == "true"
|
|
RVS_TLS_FALLBACK = os.getenv("RVS_TLS_FALLBACK", "true").lower() == "true"
|
|
RVS_TOKEN = os.getenv("RVS_TOKEN", "").strip()
|
|
|
|
LLAMA_URL = os.getenv("LLAMA_URL", "http://llama:8081").rstrip("/")
|
|
LLM_MODEL = os.getenv("LLM_MODEL", "qwen3-8b")
|
|
LLM_TIMEOUT_SEC = float(os.getenv("LLM_TIMEOUT_SEC", "60"))
|
|
|
|
# ── Compute-Fleet: Worker-Identitaet & Registrierung ──────────────
|
|
# Meldet sich bei der aria-bridge (worker_hello) + periodischer worker_ping.
|
|
NODE_NAME = os.getenv("NODE_NAME", "node").strip() or "node"
|
|
GPU_IDS = os.getenv("NVIDIA_VISIBLE_DEVICES", "").strip()
|
|
WORKER_SERVICE = "llm"
|
|
INSTANCE_ID = f"{WORKER_SERVICE}@{NODE_NAME}"
|
|
WORKER_PING_INTERVAL_S = int(os.getenv("WORKER_PING_INTERVAL_S", "10"))
|
|
_inflight = 0 # laufende llm_requests (busy-Report im ping)
|
|
# Qwen3 hat Thinking-Mode default AN — dann verbraet es Tokens in einem
|
|
# <think>-Block und liefert (bei kleinem max_tokens) leeren/abgeschnittenen
|
|
# content, ausserdem 3x langsamer. ARIAs schnelles Tier will KEIN Grübeln
|
|
# (grübeln = harter Turn = Claude). Wir schalten Thinking daher per
|
|
# chat_template_kwargs ab (Qwen3-Template versteht enable_thinking=false;
|
|
# andere Templates ignorieren das kwarg). Bei einem Modell, das darauf
|
|
# empfindlich reagiert: LLM_DISABLE_THINKING=false setzen.
|
|
LLM_DISABLE_THINKING = os.getenv("LLM_DISABLE_THINKING", "true").lower() == "true"
|
|
|
|
# ── Modell-Verwaltung (Stage D): Adapter besitzt llama-swaps Config ──
|
|
# llama-swap liest die GENERIERTE Config (beschreibbar, im /models-Bind). Wir
|
|
# erzeugen sie aus dem Basis-Template (kuratierte Defaults) + der persistenten
|
|
# Box-Registry (per Diagnostic hinzugefuegte Modelle). So werden neue Modelle
|
|
# ohne Image-Rebuild waehlbar.
|
|
import yaml # pyyaml
|
|
BASE_CONFIG_PATH = os.getenv("LLAMA_BASE_CONFIG", "/llamaswap/config.yaml")
|
|
GEN_CONFIG_PATH = os.getenv("LLAMA_GEN_CONFIG", "/models/llama-swap.config.yaml")
|
|
REGISTRY_PATH = os.getenv("LLM_REGISTRY", "/models/aria_models.json")
|
|
|
|
|
|
def _load_registry() -> list:
|
|
try:
|
|
with open(REGISTRY_PATH) as f:
|
|
data = json.load(f)
|
|
return data if isinstance(data, list) else []
|
|
except Exception:
|
|
return []
|
|
|
|
|
|
def _save_registry(reg: list) -> None:
|
|
try:
|
|
tmp = REGISTRY_PATH + ".tmp"
|
|
with open(tmp, "w") as f:
|
|
json.dump(reg, f, indent=2)
|
|
os.replace(tmp, REGISTRY_PATH)
|
|
except Exception as e:
|
|
logger.warning("Registry speichern fehlgeschlagen: %s", e)
|
|
|
|
|
|
def _generate_config() -> int:
|
|
"""Schreibt die llama-swap-Config aus Basis-Template + Registry. Gibt die
|
|
Anzahl Modelle zurueck. Idempotent, bei jeder Aenderung + beim Start."""
|
|
base = {}
|
|
try:
|
|
with open(BASE_CONFIG_PATH) as f:
|
|
base = yaml.safe_load(f) or {}
|
|
except Exception as e:
|
|
logger.warning("Basis-Template %s nicht lesbar (%s)", BASE_CONFIG_PATH, e)
|
|
models = dict(base.get("models") or {})
|
|
for e in _load_registry():
|
|
key = (e.get("key") or "").strip()
|
|
repo = (e.get("hfRepo") or "").strip()
|
|
if not key or not repo:
|
|
continue
|
|
quant = (e.get("quant") or "Q4_K_M").strip()
|
|
ctx = int(e.get("ctx") or 8192)
|
|
ngl = int(e.get("ngl") or 99)
|
|
models[key] = {
|
|
"cmd": (f"llama-server --port ${{PORT}} --host 127.0.0.1\n"
|
|
f"-hf {repo}:{quant}\n-ngl {ngl} -c {ctx} --jinja"),
|
|
"ttl": 3600,
|
|
}
|
|
out = dict(base)
|
|
out["models"] = models
|
|
try:
|
|
os.makedirs(os.path.dirname(GEN_CONFIG_PATH), exist_ok=True)
|
|
tmp = GEN_CONFIG_PATH + ".tmp"
|
|
with open(tmp, "w") as f:
|
|
yaml.safe_dump(out, f, sort_keys=False, default_flow_style=False)
|
|
os.replace(tmp, GEN_CONFIG_PATH)
|
|
logger.info("llama-swap-Config generiert: %d Modelle → %s", len(models), GEN_CONFIG_PATH)
|
|
except Exception as e:
|
|
logger.error("Config schreiben fehlgeschlagen: %s", e)
|
|
return len(models)
|
|
|
|
|
|
async def _reload_llama() -> None:
|
|
"""Stoesst llama-swap-Reload an. Viele Builds watchen die Config-Datei ohnehin;
|
|
zusaetzlich versuchen wir bekannte Reload-Endpunkte (Fehler ignoriert)."""
|
|
for path in ("/api/config/reload", "/reload"):
|
|
try:
|
|
async with httpx.AsyncClient(timeout=10) as c:
|
|
r = await c.post(f"{LLAMA_URL}{path}")
|
|
if r.status_code < 400:
|
|
logger.info("llama-swap reload via %s", path)
|
|
return
|
|
except Exception:
|
|
pass
|
|
logger.info("llama-swap reload: kein Endpoint — verlasse mich auf File-Watch")
|
|
|
|
|
|
async def _send(ws, mtype: str, payload: dict) -> None:
|
|
try:
|
|
await ws.send(json.dumps({
|
|
"type": mtype,
|
|
"payload": payload,
|
|
"timestamp": int(time.time() * 1000),
|
|
}))
|
|
except Exception as e:
|
|
logger.warning("Send fehlgeschlagen (%s): %s", mtype, e)
|
|
|
|
|
|
async def _call_llama(messages: list, *, max_tokens: int, temperature: float,
|
|
stop, tools=None, model=None) -> dict:
|
|
"""Ruft llama.cpp/llama-swap /v1/chat/completions (OpenAI-Format). Gibt
|
|
{ok, content, tool_calls, error} zurueck — wirft nie.
|
|
|
|
model: welches Modell llama-swap laden soll (B0.5). Kommt aus dem Request
|
|
(Brain -> local_llm.json). Faellt auf LLM_MODEL (env) zurueck.
|
|
tools: optionale OpenAI-Tool-Definitionen (B1b). Qwen3 (--jinja) kann
|
|
natives Tool-Calling und liefert dann message.tool_calls."""
|
|
body = {
|
|
"model": model or LLM_MODEL,
|
|
"messages": messages,
|
|
"max_tokens": max_tokens,
|
|
"temperature": temperature,
|
|
"stream": False,
|
|
}
|
|
if stop:
|
|
body["stop"] = stop
|
|
if tools:
|
|
body["tools"] = tools
|
|
body["tool_choice"] = "auto"
|
|
if LLM_DISABLE_THINKING:
|
|
# llama.cpp (--jinja) reicht chat_template_kwargs an die Chat-Vorlage
|
|
# weiter. Qwen3 unterdrueckt damit den <think>-Block.
|
|
body["chat_template_kwargs"] = {"enable_thinking": False}
|
|
try:
|
|
async with httpx.AsyncClient(timeout=LLM_TIMEOUT_SEC) as client:
|
|
r = await client.post(f"{LLAMA_URL}/v1/chat/completions", json=body)
|
|
r.raise_for_status()
|
|
data = r.json()
|
|
msg = (data.get("choices") or [{}])[0].get("message", {}) or {}
|
|
return {
|
|
"ok": True,
|
|
"content": msg.get("content") or "",
|
|
"tool_calls": msg.get("tool_calls") or None,
|
|
"usage": data.get("usage"),
|
|
}
|
|
except Exception as e:
|
|
logger.warning("llama.cpp-Call fehlgeschlagen: %s", e)
|
|
return {"ok": False, "content": "", "error": str(e)[:300]}
|
|
|
|
|
|
# B0.5-2: Lade-Status ans Diagnostic (service_status, service="llm"). Wir kennen
|
|
# den Download-Fortschritt nicht (llama-swap gibt ihn nicht her), aber wir melden
|
|
# den Zustand bei Modellwechsel: loading -> ready/error. _last_model = aktuell
|
|
# geladenes; _ready_models = in dieser Session schon einmal bereit gewesene
|
|
# (fuer den "frisch geladen"-Hinweis 🎉 bei langem Erst-Load).
|
|
_last_model = None
|
|
_ready_models: set = set()
|
|
|
|
|
|
async def _emit_llm_status(ws, state: str, model: str, **extra) -> None:
|
|
await _send(ws, "service_status",
|
|
{"service": "llm", "state": state, "model": model, **extra})
|
|
|
|
|
|
async def _fetch_available_models() -> list:
|
|
"""Fragt llama-swap ab, welche Modelle diese Box fahren kann (GET /v1/models,
|
|
OpenAI-kompatibel → {data:[{id},...]}). Das sind die config.yaml-Keys.
|
|
Defensiv: bei Fehler Fallback auf [LLM_MODEL]."""
|
|
try:
|
|
async with httpx.AsyncClient(timeout=10) as client:
|
|
r = await client.get(f"{LLAMA_URL}/v1/models")
|
|
r.raise_for_status()
|
|
data = r.json()
|
|
ids = [m.get("id") for m in (data.get("data") or []) if m.get("id")]
|
|
return ids or [LLM_MODEL]
|
|
except Exception as e:
|
|
logger.warning("llama-swap /v1/models nicht abfragbar (%s) — Fallback [%s]", e, LLM_MODEL)
|
|
return [LLM_MODEL]
|
|
|
|
|
|
async def _announce(ws) -> None:
|
|
"""Sendet ein frisches worker_hello mit der aktuellen Modell-Liste (nach
|
|
Provision/Remove aufrufen, damit Bridge+Diagnostic das neue Modell lernen)."""
|
|
models = await _fetch_available_models()
|
|
await _send(ws, "worker_hello", {
|
|
"instanceId": INSTANCE_ID, "service": WORKER_SERVICE,
|
|
"node": NODE_NAME, "gpus": GPU_IDS, "model": LLM_MODEL,
|
|
"models": models, # welche Modelle diese Box fahren kann (llama-swap-Keys)
|
|
})
|
|
logger.info("worker_hello: models=%s", models)
|
|
|
|
|
|
async def _worker_register(ws) -> None:
|
|
"""Meldet diesen Worker bei der aria-bridge an (worker_hello) und haelt die
|
|
Flotten-Registry per periodischem worker_ping (mit busy-Status) frisch."""
|
|
try:
|
|
await _announce(ws)
|
|
while True:
|
|
await asyncio.sleep(WORKER_PING_INTERVAL_S)
|
|
await _send(ws, "worker_ping",
|
|
{"instanceId": INSTANCE_ID, "busy": _inflight > 0})
|
|
except asyncio.CancelledError:
|
|
raise
|
|
except Exception:
|
|
return # Socket tot → still beenden; _run reconnectet + startet neu
|
|
|
|
|
|
async def _handle_llm_request(ws, payload: dict) -> None:
|
|
global _last_model, _inflight
|
|
_inflight += 1
|
|
try:
|
|
await _do_llm_request(ws, payload)
|
|
finally:
|
|
_inflight -= 1
|
|
|
|
|
|
async def _do_llm_request(ws, payload: dict) -> None:
|
|
global _last_model
|
|
req_id = payload.get("requestId", "")
|
|
messages = payload.get("messages") or []
|
|
if not isinstance(messages, list) or not messages:
|
|
await _send(ws, "llm_response", {
|
|
"requestId": req_id, "ok": False, "error": "leere/ungueltige messages",
|
|
})
|
|
return
|
|
max_tokens = int(payload.get("max_tokens", 512) or 512)
|
|
temperature = float(payload.get("temperature", 0.7) or 0.7)
|
|
stop = payload.get("stop")
|
|
tools = payload.get("tools") or None
|
|
model = (payload.get("model") or "").strip() or None
|
|
eff_model = model or LLM_MODEL
|
|
|
|
# Modellwechsel (oder erster Request) → llama-swap laedt/swappt: Status melden.
|
|
switching = eff_model != _last_model
|
|
if switching:
|
|
await _emit_llm_status(ws, "loading", eff_model)
|
|
|
|
t0 = time.time()
|
|
res = await _call_llama(messages, max_tokens=max_tokens,
|
|
temperature=temperature, stop=stop, tools=tools,
|
|
model=model)
|
|
dt = time.time() - t0
|
|
|
|
if switching:
|
|
if res.get("ok"):
|
|
fresh = (eff_model not in _ready_models) and dt > 25
|
|
_ready_models.add(eff_model)
|
|
_last_model = eff_model
|
|
await _emit_llm_status(ws, "ready", eff_model,
|
|
loadSeconds=round(dt, 1), freshlyDownloaded=fresh)
|
|
else:
|
|
# bei Fehler _last_model NICHT setzen → naechster Versuch meldet erneut loading
|
|
await _emit_llm_status(ws, "error", eff_model,
|
|
error=(res.get("error") or "")[:120])
|
|
tc = res.get("tool_calls")
|
|
logger.info("llm_request id=%s model=%s -> ok=%s %.2fs content_len=%d tool_calls=%d",
|
|
(req_id[:8] if req_id else "?"), model or LLM_MODEL, res.get("ok"), dt,
|
|
len(res.get("content") or ""), len(tc) if tc else 0)
|
|
await _send(ws, "llm_response", {
|
|
"requestId": req_id,
|
|
"ok": res.get("ok", False),
|
|
"content": res.get("content", ""),
|
|
"tool_calls": tc,
|
|
"error": res.get("error"),
|
|
"model": model or LLM_MODEL,
|
|
"elapsedMs": int(dt * 1000),
|
|
})
|
|
|
|
|
|
async def _handle_provision(ws, payload: dict) -> None:
|
|
"""Fuegt ein Modell hinzu: Registry+Config schreiben, reload, dann Warmup
|
|
(zieht das GGUF via -hf beim ersten Load). Meldet die neue Modell-Liste."""
|
|
key = (payload.get("key") or "").strip()
|
|
repo = (payload.get("hfRepo") or "").strip()
|
|
if not key or not repo:
|
|
await _send(ws, "llm_provision_result",
|
|
{"instanceId": INSTANCE_ID, "key": key, "ok": False, "error": "key/hfRepo fehlt"})
|
|
return
|
|
entry = {
|
|
"key": key, "hfRepo": repo,
|
|
"quant": (payload.get("quant") or "Q4_K_M").strip(),
|
|
"ctx": int(payload.get("ctx") or 8192),
|
|
"ngl": int(payload.get("ngl") or 99),
|
|
}
|
|
reg = [e for e in _load_registry() if e.get("key") != key]
|
|
reg.append(entry)
|
|
_save_registry(reg)
|
|
_generate_config()
|
|
await _reload_llama()
|
|
await _announce(ws) # Bridge/Diagnostic lernen das neue Modell
|
|
# Warmup: Mini-Request → llama-swap laedt/zieht das Modell (Fortschritt via
|
|
# service_status loading→ready, freshlyDownloaded).
|
|
await _emit_llm_status(ws, "loading", key)
|
|
t0 = time.time()
|
|
res = await _call_llama([{"role": "user", "content": "hi"}],
|
|
max_tokens=1, temperature=0.0, stop=None, model=key)
|
|
dt = time.time() - t0
|
|
if res.get("ok"):
|
|
_ready_models.add(key)
|
|
await _emit_llm_status(ws, "ready", key, loadSeconds=round(dt, 1),
|
|
freshlyDownloaded=dt > 25)
|
|
else:
|
|
await _emit_llm_status(ws, "error", key, error=(res.get("error") or "")[:160])
|
|
await _send(ws, "llm_provision_result",
|
|
{"instanceId": INSTANCE_ID, "key": key, "ok": res.get("ok", False),
|
|
"error": res.get("error"), "elapsedMs": int(dt * 1000)})
|
|
logger.info("provision %s (%s) → ok=%s %.1fs", key, repo, res.get("ok"), dt)
|
|
|
|
|
|
async def _handle_remove(ws, payload: dict) -> None:
|
|
"""Entfernt ein Modell aus Registry+Config (GGUF bleibt im Cache)."""
|
|
key = (payload.get("key") or "").strip()
|
|
if not key:
|
|
return
|
|
reg = [e for e in _load_registry() if e.get("key") != key]
|
|
_save_registry(reg)
|
|
_generate_config()
|
|
await _reload_llama()
|
|
await _announce(ws)
|
|
await _send(ws, "llm_provision_result",
|
|
{"instanceId": INSTANCE_ID, "key": key, "ok": True, "removed": True})
|
|
logger.info("removed model %s", key)
|
|
|
|
|
|
async def _run() -> None:
|
|
if not RVS_HOST:
|
|
logger.error("RVS_HOST nicht gesetzt — Abbruch")
|
|
return
|
|
if not RVS_TOKEN:
|
|
logger.error("RVS_TOKEN nicht gesetzt — Abbruch")
|
|
return
|
|
|
|
# llama-swap-Config aus Basis-Template + Registry erzeugen, BEVOR llama-swap
|
|
# sie braucht (llama-swap restart: unless-stopped faengt die Erst-Boot-
|
|
# Reihenfolge ab, falls es kurz vor uns startet).
|
|
_generate_config()
|
|
|
|
use_tls = RVS_TLS
|
|
retry_s = 2
|
|
tls_fallback_tried = False
|
|
|
|
while True:
|
|
scheme = "wss" if use_tls else "ws"
|
|
url = f"{scheme}://{RVS_HOST}:{RVS_PORT}/ws?token={RVS_TOKEN}"
|
|
masked = url.replace(RVS_TOKEN, "***") if RVS_TOKEN else url
|
|
try:
|
|
logger.info("Verbinde zu RVS: %s (llama=%s)", masked, LLAMA_URL)
|
|
async with websockets.connect(
|
|
url, ping_interval=20, ping_timeout=10, max_size=16 * 1024 * 1024
|
|
) as ws:
|
|
logger.info("RVS verbunden — llm-adapter online")
|
|
retry_s = 2
|
|
tls_fallback_tried = False
|
|
ping_task = asyncio.create_task(_worker_register(ws))
|
|
async for raw in ws:
|
|
try:
|
|
msg = json.loads(raw)
|
|
except Exception:
|
|
continue
|
|
mtype = msg.get("type")
|
|
if mtype not in ("llm_request", "llm_provision_model", "llm_remove_model"):
|
|
continue
|
|
payload = msg.get("payload", {}) or {}
|
|
# Redundanz-Routing: gezielt an eine andere Instanz adressiert
|
|
# → ignorieren. Ohne targetInstance → wie bisher (jeder nimmt).
|
|
tgt = payload.get("targetInstance")
|
|
if tgt and tgt != INSTANCE_ID:
|
|
continue
|
|
if mtype == "llm_provision_model":
|
|
asyncio.create_task(_handle_provision(ws, payload))
|
|
elif mtype == "llm_remove_model":
|
|
asyncio.create_task(_handle_remove(ws, payload))
|
|
else:
|
|
# Jede Anfrage nebenlaeufig — llama.cpp serialisiert intern,
|
|
# aber wir blockieren so nicht den Empfang weiterer Messages.
|
|
asyncio.create_task(_handle_llm_request(ws, payload))
|
|
except Exception as e:
|
|
logger.warning("RVS-Verbindung verloren/fehlgeschlagen: %s", e)
|
|
try:
|
|
ping_task.cancel()
|
|
except NameError:
|
|
pass
|
|
if use_tls and RVS_TLS_FALLBACK and not tls_fallback_tried:
|
|
tls_fallback_tried = True
|
|
use_tls = False
|
|
logger.info("TLS fehlgeschlagen — Fallback auf ws://")
|
|
continue
|
|
await asyncio.sleep(min(retry_s, 30))
|
|
retry_s = min(retry_s * 2, 30)
|
|
use_tls = RVS_TLS
|
|
|
|
|
|
if __name__ == "__main__":
|
|
asyncio.run(_run())
|