Update app.py

app.py

@@ -21,9 +21,9 @@ ROLE_TOKENS = {
     "system": SYSTEM_TOKEN
 }
 
-CONTEXT_SIZE = 4000
+CONTEXT_SIZE = 2000
 ENABLE_GPU = True
-GPU_LAYERS = 70
+GPU_LAYERS = 20
 
 # Create a lock object
 lock = threading.Lock()