Update app.py

app.py

@@ -119,7 +119,6 @@ with gr.Blocks() as demo:
 
     def user_fn(user_message, history: list):
         return "", history + [{"role": "user", "content": user_message}]
-
     
     def generate_fn(history: list):
          # continue from user input
@@ -131,17 +130,17 @@ with gr.Blocks() as demo:
             
         history.append({"role": "assistant", "content": ""})
         # history[-1]["content"] += "Generating with the given prefix...\n"
-        for history in model_fn(prefix):
+        for history in model_fn(prefix, history):
             yield history
 
     def continue_fn(history: list):
          # continue from the last model output
         prefix = history[-1]["content"]
-        for history in model_fn(prefix):
+        for history in model_fn(prefix, history):
             yield history
         
     @spaces.GPU
-    def model_fn(prefix):
+    def model_fn(prefix, history):
         queue = Queue(maxsize=10)
         class MyStreamer:
             def put(self, tokens):