feat(memory): Ajout déchargement automatique mémoire sur FolzoCorp
- effi-stt: Déchargement Whisper après 300s inactivité (INACTIVITY_TIMEOUT) * Nouveaux endpoints: /unload, /reload, /stats * Modelo rechargé à la première requête - ComfyUI: Script monitor qui décharge VRAM après inactivité * Vérifie usage VRAM toutes les 60s * Exécute workflow UnloadAllModelsNode après 300s inactivité - Kreuzberg: Manager qui redémarre le container après seuil mémoire + inactivité * Restart si > 2GB pendant > 10min inactif - Documentation: docs/MEMOIRE-GESTION.md
This commit is contained in:
Folzo
Executable
+182
@@ -0,0 +1,182 @@
|
||||
#!/usr/bin/env python3
|
||||
# ComfyUI Model Unloader - Décharge les modèles après inactivité
|
||||
# Usage: python3 comfyui-unload-monitor.py [--interval 60] [--timeout 300]
|
||||
|
||||
import requests
|
||||
import time
|
||||
import argparse
|
||||
import logging
|
||||
import sys
|
||||
import signal
|
||||
import subprocess
|
||||
from datetime import datetime, timedelta
|
||||
|
||||
logging.basicConfig(
|
||||
level=logging.INFO,
|
||||
format='%(asctime)s - %(levelname)s - %(message)s'
|
||||
)
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
DEFAULT_INTERVAL = 60
|
||||
DEFAULT_TIMEOUT = 300
|
||||
COMFYUI_URL = "http://localhost:8188"
|
||||
|
||||
class ComfyUIUnloadMonitor:
|
||||
def __init__(self, interval=DEFAULT_INTERVAL, timeout=DEFAULT_TIMEOUT):
|
||||
self.interval = interval
|
||||
self.timeout = timeout
|
||||
self.last_activity = datetime.now()
|
||||
self.running = True
|
||||
self.models_loaded = False
|
||||
|
||||
signal.signal(signal.SIGINT, self.signal_handler)
|
||||
signal.signal(signal.SIGTERM, self.signal_handler)
|
||||
|
||||
def signal_handler(self, signum, frame):
|
||||
logger.info("Signal reçu, arrêt du monitor...")
|
||||
self.running = False
|
||||
|
||||
def check_comfyui_activity(self):
|
||||
try:
|
||||
response = requests.get(f"{COMFYUI_URL}/system_stats", timeout=5)
|
||||
if response.status_code == 200:
|
||||
stats = response.json()
|
||||
return stats
|
||||
return None
|
||||
except requests.exceptions.RequestException as e:
|
||||
logger.warning(f"ComfyUI non accessible: {e}")
|
||||
return None
|
||||
|
||||
def check_queue_and_history(self):
|
||||
try:
|
||||
queue_response = requests.get(f"{COMFYUI_URL}/queue", timeout=5)
|
||||
history_response = requests.get(f"{COMFYUI_URL}/history?max_items=1", timeout=5)
|
||||
|
||||
if queue_response.status_code == 200 and history_response.status_code == 200:
|
||||
queue_data = queue_response.json()
|
||||
history_data = history_response.json()
|
||||
|
||||
running = queue_data.get("queue_running", [])
|
||||
pending = queue_data.get("queue_pending", [])
|
||||
|
||||
has_activity = len(running) > 0 or len(pending) > 0 or len(history_data) > 0
|
||||
|
||||
if history_data:
|
||||
latest = list(history_data.values())[0]
|
||||
if "nodes" in latest:
|
||||
self.last_activity = datetime.now()
|
||||
logger.debug(f"Activité détectée: {len(running)} en cours, {len(pending)} en attente")
|
||||
|
||||
return has_activity
|
||||
return False
|
||||
except Exception as e:
|
||||
logger.debug(f"Erreur check_queue: {e}")
|
||||
return False
|
||||
|
||||
def get_memory_usage(self):
|
||||
stats = self.check_comfyui_activity()
|
||||
if stats and "memory" in stats:
|
||||
vram_used = stats["memory"].get("vram_used", 0)
|
||||
ram_used = stats["memory"].get("ram_used", 0)
|
||||
total_vram = stats["memory"].get("vram_total", 0)
|
||||
|
||||
if total_vram > 0:
|
||||
vram_percent = (vram_used / total_vram) * 100
|
||||
return vram_used, ram_used, vram_percent
|
||||
return None, None, None
|
||||
|
||||
def unload_all_models(self):
|
||||
logger.info("🔄 Déchargement de tous les modèles ComfyUI...")
|
||||
|
||||
workflow = {
|
||||
"3": {
|
||||
"inputs": {"value": "unload_trigger"},
|
||||
"class_type": "UnloadAllModelsNode"
|
||||
}
|
||||
}
|
||||
|
||||
try:
|
||||
response = requests.post(
|
||||
f"{COMFYUI_URL}/prompt",
|
||||
json={"prompt": workflow},
|
||||
timeout=30
|
||||
)
|
||||
|
||||
if response.status_code == 200:
|
||||
logger.info("✅ Modèles déchargés avec succès")
|
||||
self.models_loaded = False
|
||||
return True
|
||||
else:
|
||||
logger.error(f"❌ Erreur déchargement: {response.status_code}")
|
||||
return False
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"❌ Exception lors du déchargement: {e}")
|
||||
return False
|
||||
|
||||
def soft_cleanup(self):
|
||||
logger.info("🧹 Nettoyage soft (gc.collect + empty cache)...")
|
||||
try:
|
||||
requests.post(f"{COMFYUI_URL}/internal/cleanup", timeout=10)
|
||||
except:
|
||||
pass
|
||||
|
||||
vram, ram, percent = self.get_memory_usage()
|
||||
if vram is not None:
|
||||
logger.info(f"📊 VRAM après cleanup: {vram/1024**3:.2f}GB ({percent:.1f}%)")
|
||||
|
||||
def run(self):
|
||||
logger.info(f"🚀 Démarrage ComfyUI Unload Monitor")
|
||||
logger.info(f" Intervalle de vérification: {self.interval}s")
|
||||
logger.info(f" Timeout d'inactivité: {self.timeout}s")
|
||||
logger.info(f" URL ComfyUI: {COMFYUI_URL}")
|
||||
|
||||
while self.running:
|
||||
try:
|
||||
vram, ram, vram_percent = self.get_memory_usage()
|
||||
|
||||
if vram is not None and vram_percent > 1:
|
||||
self.models_loaded = True
|
||||
idle_time = (datetime.now() - self.last_activity).total_seconds()
|
||||
|
||||
logger.debug(f"📊 VRAM: {vram/1024**3:.2f}GB ({vram_percent:.1f}%), "
|
||||
f"RAM: {ram/1024**3:.2f}GB, "
|
||||
f"Inactivité: {idle_time:.0f}s")
|
||||
|
||||
if idle_time >= self.timeout:
|
||||
logger.info(f"⏰ Inactivité de {idle_time:.0f}s >= {self.timeout}s")
|
||||
self.unload_all_models()
|
||||
time.sleep(5)
|
||||
self.soft_cleanup()
|
||||
self.last_activity = datetime.now()
|
||||
else:
|
||||
if self.models_loaded:
|
||||
logger.info("📊 Modèles semblent déchargés (VRAM < 1%)")
|
||||
self.models_loaded = False
|
||||
|
||||
time.sleep(self.interval)
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Erreur dans la boucle principale: {e}")
|
||||
time.sleep(self.interval)
|
||||
|
||||
logger.info("👋 Monitor arrêté")
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="ComfyUI Model Unload Monitor")
|
||||
parser.add_argument("--interval", type=int, default=DEFAULT_INTERVAL,
|
||||
help=f"Intervalle de vérification en secondes (défaut: {DEFAULT_INTERVAL})")
|
||||
parser.add_argument("--timeout", type=int, default=DEFAULT_TIMEOUT,
|
||||
help=f"Timeout d'inactivité en secondes (défaut: {DEFAULT_TIMEOUT})")
|
||||
parser.add_argument("--debug", action="store_true", help="Mode debug")
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.debug:
|
||||
logging.getLogger().setLevel(logging.DEBUG)
|
||||
|
||||
monitor = ComfyUIUnloadMonitor(interval=args.interval, timeout=args.timeout)
|
||||
monitor.run()
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Executable
+174
@@ -0,0 +1,174 @@
|
||||
#!/usr/bin/env python3
|
||||
# Kreuzberg Memory Manager - Redémarre le container après inactivité
|
||||
# Usage: python3 kreuzberg-memory-manager.py [--interval 60] [--timeout 600] [--max-memory 2]
|
||||
|
||||
import requests
|
||||
import time
|
||||
import argparse
|
||||
import logging
|
||||
import subprocess
|
||||
import sys
|
||||
|
||||
logging.basicConfig(
|
||||
level=logging.INFO,
|
||||
format='%(asctime)s - %(levelname)s - %(message)s'
|
||||
)
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
DEFAULT_INTERVAL = 60
|
||||
DEFAULT_TIMEOUT = 600
|
||||
DEFAULT_MAX_MEMORY = 2.0
|
||||
KREUZBERG_API = "http://localhost:8300"
|
||||
CONTAINER_NAME = "kreuzberg"
|
||||
|
||||
class KreuzbergMemoryManager:
|
||||
def __init__(self, interval=DEFAULT_INTERVAL, timeout=DEFAULT_TIMEOUT, max_memory_gb=DEFAULT_MAX_MEMORY):
|
||||
self.interval = interval
|
||||
self.timeout = timeout
|
||||
self.max_memory_gb = max_memory_gb
|
||||
self.last_activity = time.time()
|
||||
self.running = True
|
||||
self.container_id = None
|
||||
|
||||
import signal
|
||||
signal.signal(signal.SIGINT, self.signal_handler)
|
||||
signal.signal(signal.SIGTERM, self.signal_handler)
|
||||
|
||||
self.find_container()
|
||||
|
||||
def signal_handler(self, signum, frame):
|
||||
logger.info("Signal reçu, arrêt du manager...")
|
||||
self.running = False
|
||||
|
||||
def find_container(self):
|
||||
try:
|
||||
result = subprocess.run(
|
||||
["docker", "ps", "-q", "--filter", f"name={CONTAINER_NAME}"],
|
||||
capture_output=True, text=True
|
||||
)
|
||||
if result.stdout.strip():
|
||||
self.container_id = result.stdout.strip()
|
||||
logger.info(f"Container trouvé: {self.container_id[:12]}")
|
||||
else:
|
||||
logger.warning(f"Container {CONTAINER_NAME} non trouvé")
|
||||
except Exception as e:
|
||||
logger.error(f"Erreur recherche container: {e}")
|
||||
|
||||
def get_container_memory(self):
|
||||
if not self.container_id:
|
||||
return None
|
||||
try:
|
||||
result = subprocess.run(
|
||||
["docker", "stats", "--no-stream", "--format", "{{.MemUsage}}", self.container_id],
|
||||
capture_output=True, text=True
|
||||
)
|
||||
if result.stdout.strip():
|
||||
mem_str = result.stdout.strip()
|
||||
return self.parse_memory(mem_str)
|
||||
except Exception as e:
|
||||
logger.debug(f"Erreur stats: {e}")
|
||||
return None
|
||||
|
||||
def parse_memory(self, mem_str):
|
||||
try:
|
||||
parts = mem_str.split("/")
|
||||
if len(parts) >= 1:
|
||||
used = parts[0].strip()
|
||||
if "GiB" in used:
|
||||
return float(used.replace("GiB", "")) * 1024
|
||||
elif "MiB" in used:
|
||||
return float(used.replace("MiB", ""))
|
||||
elif "GB" in used:
|
||||
return float(used.replace("GB", "")) * 1024
|
||||
elif "MB" in used:
|
||||
return float(used.replace("MB", ""))
|
||||
except:
|
||||
pass
|
||||
return None
|
||||
|
||||
def check_api_activity(self):
|
||||
try:
|
||||
response = requests.get(f"{KREUZBERG_API}/health", timeout=5)
|
||||
if response.status_code in [200, 404]:
|
||||
self.last_activity = time.time()
|
||||
return True
|
||||
except:
|
||||
pass
|
||||
return False
|
||||
|
||||
def restart_container(self):
|
||||
logger.info(f"🔄 Redémarrage du container {CONTAINER_NAME}...")
|
||||
try:
|
||||
subprocess.run(["docker", "restart", CONTAINER_NAME], check=True)
|
||||
logger.info("✅ Container redémarré")
|
||||
time.sleep(5)
|
||||
self.find_container()
|
||||
return True
|
||||
except Exception as e:
|
||||
logger.error(f"❌ Erreur redémarrage: {e}")
|
||||
return False
|
||||
|
||||
def run(self):
|
||||
logger.info(f"🚀 Démarrage Kreuzberg Memory Manager")
|
||||
logger.info(f" Intervalle: {self.interval}s")
|
||||
logger.info(f" Timeout inactivité: {self.timeout}s")
|
||||
logger.info(f" Seuil mémoire: {self.max_memory_gb}GB")
|
||||
|
||||
idle_time = 0
|
||||
|
||||
while self.running:
|
||||
try:
|
||||
memory_mb = self.get_container_memory()
|
||||
|
||||
if memory_mb is not None:
|
||||
memory_gb = memory_mb / 1024
|
||||
|
||||
is_active = self.check_api_activity()
|
||||
|
||||
if is_active:
|
||||
idle_time = 0
|
||||
self.last_activity = time.time()
|
||||
logger.debug(f"📊 Mémoire: {memory_gb:.2f}GB, API active")
|
||||
else:
|
||||
idle_time += self.interval
|
||||
logger.debug(f"📊 Mémoire: {memory_gb:.2f}GB, Inactif: {idle_time}s")
|
||||
|
||||
if memory_gb > self.max_memory_gb:
|
||||
logger.warning(f"⚠️ Mémoire ({memory_gb:.2f}GB) > seuil ({self.max_memory_gb}GB)")
|
||||
if idle_time >= self.timeout:
|
||||
logger.info(f"⏰ Seuil timeout atteint ({idle_time}s >= {self.timeout}s)")
|
||||
self.restart_container()
|
||||
idle_time = 0
|
||||
|
||||
time.sleep(self.interval)
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Erreur: {e}")
|
||||
time.sleep(self.interval)
|
||||
|
||||
logger.info("👋 Manager arrêté")
|
||||
|
||||
def main():
|
||||
parser = argparse.ArgumentParser(description="Kreuzberg Memory Manager")
|
||||
parser.add_argument("--interval", type=int, default=DEFAULT_INTERVAL,
|
||||
help=f"Intervalle (défaut: {DEFAULT_INTERVAL}s)")
|
||||
parser.add_argument("--timeout", type=int, default=DEFAULT_TIMEOUT,
|
||||
help=f"Timeout inactivité (défaut: {DEFAULT_TIMEOUT}s)")
|
||||
parser.add_argument("--max-memory", type=float, default=DEFAULT_MAX_MEMORY,
|
||||
help=f"Seuil mémoire GB (défaut: {DEFAULT_MAX_MEMORY}GB)")
|
||||
parser.add_argument("--debug", action="store_true", help="Mode debug")
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
if args.debug:
|
||||
logging.getLogger().setLevel(logging.DEBUG)
|
||||
|
||||
manager = KreuzbergMemoryManager(
|
||||
interval=args.interval,
|
||||
timeout=args.timeout,
|
||||
max_memory_gb=args.max_memory
|
||||
)
|
||||
manager.run()
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user