Update app.py

app.py

@@ -88,17 +88,19 @@ def normalize_history(history):
             continue
     return normalized
 
-# 在构建 tokens 之前调用
-history = normalize_history(history)
 
 # ---------------------------------------------------------------------------
 # Chat function (streaming)
 # ---------------------------------------------------------------------------
 def chat(message, history, temperature, max_tokens, top_k):
     """Generate a streaming response for the chat interface."""
+    # 在构建 tokens 之前调用
+    history = normalize_history(history)
     # Build the full token sequence from conversation history
     # history is a list of [user_msg, assistant_msg] tuples
     tokens = [bos_id]
+
+
     for user_msg, assistant_msg in history:
         tokens.append(user_start)
         tokens.extend(tokenizer.encode(user_msg))