fix model init

app.py

@@ -367,7 +367,11 @@ def chat(system_prompt, user_text, image_file, history, max_tokens, temperature,
 
     start_time = time.time()
     try:
-        model, processor = get_model(model_name)
+        # 直接复用已加载的全局模型，必要时再触发加载
+        model = _MODEL
+        processor = _PROCESSOR
+        if model is None or processor is None or _LOADED_MODEL_NAME != model_name:
+            model, processor = get_model(model_name)
 
         inputs = processor.apply_chat_template(
             messages, add_generation_prompt=True, tokenize=True,