diff --git a/vllm_ascend/platform.py b/vllm_ascend/platform.py
index 65db170..45647b0 100644
--- a/vllm_ascend/platform.py
+++ b/vllm_ascend/platform.py
@@ -103,9 +103,6 @@ class NPUPlatform(Platform):
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
-        # Register ops when setup.
-        from vllm_ascend import ops  # noqa: F401
-
         parallel_config = vllm_config.parallel_config
         if parallel_config.worker_cls == "auto":
             parallel_config.worker_cls = "vllm_ascend.worker.NPUWorker"
diff --git a/vllm_ascend/worker.py b/vllm_ascend/worker.py
index c5884e3..27930d4 100644
--- a/vllm_ascend/worker.py
+++ b/vllm_ascend/worker.py
@@ -68,6 +68,8 @@ class NPUWorker(LocalOrDistributedWorkerBase):
         is_driver_worker: bool = False,
         model_runner_cls: Optional[Type[ModelRunnerBase]] = None,
     ) -> None:
+        # Register ops when worker init.
+        from vllm_ascend import ops  # noqa: F401
 
         WorkerBase.__init__(self, vllm_config=vllm_config)
         # Try to import mindie_turbo to accelerate vLLM inference.