Update app.py

app.py

@@ -6,6 +6,7 @@ from llama_cpp import Llama
 import requests
 import tempfile
 import json
+from concurrent.futures import ThreadPoolExecutor
 
 app = Flask(__name__)
 logging.basicConfig(level=logging.INFO)
@@ -20,6 +21,7 @@ class LLMManager:
     def __init__(self, models_config):
         self.models = {}
         self.models_config = models_config
+        self.executor = ThreadPoolExecutor(max_workers=2)
         self.load_all_models()
 
     def load_all_models(self):