Upload app.py

app.py

@@ -141,7 +141,7 @@ def merlin_chat(message, history):
     prompt = "[" + json.dumps(system_msg) + chat_json + json.dumps(user_msg) + "{'role': 'assistant, 'content': '*recalls \""
     for attempt in range(3):
         # result = model(prompt, max_new_tokens=250, return_full_text=False, handle_long_generation="hole")
-        result = generate_text(prompt, model_path, parameters)
+        result = generate_text(prompt, model_path, parameters, headers)
         response = result[0]['generated_text']
         start = 0
         end = 0