Update app.py

app.py

@@ -28,7 +28,7 @@ class ChatClient:
         self.model = AutoModelForCausalLM.from_pretrained(model_path).to(self.device)
         self.model.eval()  # 设置为评估模式
 
-    def chat_completion(self, messages, max_tokens, stream=False, temperature=1.0, top_p=1.0):
+    async def chat_completion(self, messages, max_tokens, stream=False, temperature=1.0, top_p=1.0):
         """
         生成对话回复。
         """
@@ -56,6 +56,7 @@ class ChatClient:
         # 解码生成的文本
         for sequence in output_sequences:
             result_text = self.tokenizer.decode(sequence, skip_special_tokens=True)
+            await anyio.sleep(0)  # Yield control, simulating asynchronous operation
             yield result_text
 
 # 创建客户端实例，指定模型路径