diff --git a/vllm_npu/utils.py b/vllm_npu/utils.py
index a2d7909..3f3d2e1 100644
--- a/vllm_npu/utils.py
+++ b/vllm_npu/utils.py
@@ -728,9 +728,14 @@ def weak_ref_tensor(tensor: Any) -> Any:
     Create a weak reference to a tensor.
     The new tensor will share the same data as the original tensor,
     but will not keep the original tensor alive.
+    Falls back to returning the tensor as-is if the C++ op is unavailable.
     """
     if isinstance(tensor, torch.Tensor):
-        return torch.ops._C_ascend.weak_ref_tensor(tensor)
+        if hasattr(torch.ops, '_C_ascend') and hasattr(
+                torch.ops._C_ascend, 'weak_ref_tensor'):
+            return torch.ops._C_ascend.weak_ref_tensor(tensor)
+        # Fallback: return tensor directly (no weak ref, minor memory overhead)
+        return tensor
     else:
         return tensor