Update app.py

app.py

@@ -138,7 +138,7 @@ def generate_response_non_streaming(instruction, model_name, temperature=0.7, ma
         )
         
         # 解码并返回生成的文本
-        generated_text = current_tokenizer.decode(output[0], skip_special_tokens=True)#[len(chat_input[0]):]
+        generated_text = current_tokenizer.decode(output[0][len(chat_input[0]):], skip_special_tokens=True)
         print("generated_text:",generated_text)
         return generated_text
     except Exception as e: