changes

app.py

@@ -29,7 +29,7 @@ def generate_text(input_text, history):
     else:
         input_text_with_history = f"{history[-1][1]}" + "\n"
         input_text_with_history += f"USER: {input_text}" + "\n" + " ASSISTANT:"
-    output = llm(input_text_with_history, max_tokens=1024, stop=[
+    output = llm(input_text_with_history, max_tokens=4096, stop=[
         "<|prompter|>", "<|endoftext|>", "<|endoftext|> \n",
         "ASSISTANT:", "USER:", "SYSTEM:"], stream=True
     )
@@ -52,7 +52,6 @@ demo = gr.ChatInterface(generate_text,
                                   "Ты говоришь по русски? Я злой."
                                   ],
                         cache_examples=True,
-                        retry_btn="Retry",
                         undo_btn="Undo",
                         clear_btn="Clear")