Update app.py

app.py

@@ -60,13 +60,14 @@ Age=""
 chat_log_name =""
 
 from llama_cpp import Llama
+llm = Llama(model_path=model_file, model_type="mistral",n_gpu_layers=-1,n_ctx = 2048)
+
 
 def load_model():
+    global llm
     llm = Llama(model_path=model_file, model_type="mistral",n_gpu_layers=-1,n_ctx = 2048)
     return "Model loaded"
 
-load_model()
-
 def reset_chat():
     global chat_history, chat_log_history, isFirstRun, context, chat_log_name
     chat_history = []