Update app.py

app.py

@@ -3,6 +3,9 @@ import os
 import logging
 import threading
 import time
+import subprocess
+import sys
+subprocess.check_call([sys.executable, "-m", "pip", "install", "llama-cpp-python==0.3.15"])
 from llama_cpp import Llama
 import requests
 import tempfile
@@ -14,7 +17,7 @@ app = Flask(__name__)
 logging.basicConfig(level=logging.INFO)
 
 MAX_CONTEXT_TOKENS = 1024 * 8
-MAX_GENERATION_TOKENS = 1024 * 4
+MAX_GENERATION_TOKENS = 1024 * 8
 
 with open('engines.json', 'r') as f:
     MODELS = json.load(f)