diff --git a/vllm_ascend/envs.py b/vllm_ascend/envs.py
index f78c856..8d7c10e 100644
--- a/vllm_ascend/envs.py
+++ b/vllm_ascend/envs.py
@@ -127,7 +127,7 @@ env_variables: Dict[str, Callable[[], Any]] = {
     # We set this var default to `1` in vllm-ascend to avoid segment fault when
     # enable `pin_memory` while creating a tensor using `torch.tensor`.
     "VLLM_ASCEND_ACL_OP_INIT_MODE":
-    lambda: os.getenv("VLLM_ASCEND_ACL_OP_INIT_MODE", '1'),
+    lambda: os.getenv("VLLM_ASCEND_ACL_OP_INIT_MODE", '0'),
     # Some models are optimized by vllm ascend. While in some case, e.g. rlhf
     # training, the optimized model may not be suitable. In this case, set this
     # value to False to disable the optimized model.