Update app.py

app.py

@@ -30,15 +30,15 @@ with gr.Blocks(theme=theme) as demo:
         instruction = history[-1][1] or ""
         user_message = history[-1][0]
         
-        # token_instruction_header = b"### Instruction: "
-        # token_instruction_text = instruction.encode()
+        token_instruction_header = b"### Instruction: "
+        token_instruction_text = instruction.encode()
         
-        # token_user_header = b"\n\n### User: "
+        token_user_header = b"\n\n### User: "
         token_user_text = user_message.encode()
         
         token_response_header = b"\n\n### Response:"
         
-        tokens = llm2.tokenize(token_user_text + token_response_header)
+        tokens = llm2.tokenize(token_instruction_header + token_instruction_text + token_user_header + token_user_text + token_response_header)
         
         history[-1][1] = ""
         count = 0