Update app.py

app.py

@@ -554,6 +554,7 @@ def chat_fn(message, history, temperature, top_k, top_p, max_tokens):
     conversation += f"User: {message} Sam:"
     
     # Stream response token by token
+    partial_response = ""
     for response in model.generate_stream(
         conversation,
         max_new_tokens=max_tokens,
@@ -561,7 +562,9 @@ def chat_fn(message, history, temperature, top_k, top_p, max_tokens):
         top_k=top_k,
         top_p=top_p
     ):
-        yield response
+        partial_response = response
+        # Yield the full history + current streaming message
+        yield history + [[message, partial_response]]
 
 # Create Gradio interface
 with gr.Blocks(theme=gr.themes.Soft()) as demo: