Update app.py

app.py

@@ -54,9 +54,11 @@ def respond(
     prompt = "<s>"
     
     for m in history:
-        prompt += f"{user_name}: {m[0]}\n{ai_name}: {m[1]}\n"
-    
-    for response in generate(prompt, use_rpc=use_rpc, max_tokens=max_tokens)
+        prompt += f"{user_name}: {m[0]}\n{ai_name}: {m[1]}"
+
+    response = ""
+    for tok in generate(prompt, use_rpc=use_rpc, max_tokens=max_tokens)
+        respond += tok
         yield response