Update app.py

app.py

@@ -25,6 +25,7 @@ def generate(
     temperature: float = 0.7,
     top_p: float = 1.0,
     repetition_penalty: float = 1.1,
+    stop_ids: list[list[int]] = [[2, 6, 7, 8]]
 ) -> Iterator[str]:
     conversation = []
     if system_prompt:
@@ -49,6 +50,7 @@ def generate(
         temperature=temperature,
         num_beams=1,
         repetition_penalty=repetition_penalty,
+        stop_ids=stop_ids
     )
     t = Thread(target=model.generate, kwargs=generate_kwargs)
     t.start()
@@ -61,7 +63,7 @@ def generate(
 chat_interface = gr.ChatInterface(
     fn=generate,
     additional_inputs=[
-        gr.Textbox(label="System prompt", lines=6),
+        gr.Textbox(label="System prompt", lines=8),  # Increase lines for larger box
         gr.Slider(
             label="Max new tokens",
             minimum=1,
@@ -91,7 +93,7 @@ chat_interface = gr.ChatInterface(
             value=1.1,
         ),
     ],
-    stop_btn=None,
+    stop_token=None,  # Remove stop token
     examples=[
         ["Hello there! How are you doing?"],
         ["Can you explain briefly to me what is the Python programming language?"],
@@ -105,4 +107,4 @@ with gr.Blocks(css="style.css") as demo:
     chat_interface.render()
 
 if __name__ == "__main__":
-    demo.queue(max_size=20).launch()
+    demo.queue(max_size=20).launch()