diff --git a/src/twinkle/kernel/builtin.py b/src/twinkle/kernel/builtin.py
index faffb4e7..7675767b 100644
--- a/src/twinkle/kernel/builtin.py
+++ b/src/twinkle/kernel/builtin.py
@@ -90,10 +90,11 @@ def npu_builtin(model: nn.Module | None = None) -> dict[Any, dict[str, Any]]:
         if exists('mindspeed>=0.12.1') and exists('flash-linear-attention'):
             apply_qwen3_5_fla(model)
         else:
-            logger.warning('[NPU] [FLA] mindspeed or flash-linear-attention is not installed, or mindspeed version smaller than 0.12.1; '
-                           'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
-                           'Install them with: pip install flash-linear-attention and mindspeed from source code via https://gitcode.com/Ascend/MindSpeed.'
-                           'Other NPU patches will still apply.')
+            logger.warning(
+                '[NPU] [FLA] mindspeed or flash-linear-attention is not installed, or mindspeed version smaller than 0.12.1; '
+                'FLA patch for Qwen3.5 requires mindspeed and flash-linear-attention. '
+                'Install them with: pip install flash-linear-attention and mindspeed from source code via https://gitcode.com/Ascend/MindSpeed.'
+                'Other NPU patches will still apply.')
     return bundle
 
 
diff --git a/src/twinkle/model/transformers/transformers.py b/src/twinkle/model/transformers/transformers.py
index c8fec86d..9be48cae 100644
--- a/src/twinkle/model/transformers/transformers.py
+++ b/src/twinkle/model/transformers/transformers.py
@@ -384,7 +384,7 @@ def _router_replay_setup(self, router_replay_action, routed_experts=None, batch_
         Returns ``cleanup_fn``.
         Call *cleanup_fn* after the forward to gather recorded indices(when RECORD) and clear global state.
         """
-        if not self._router_replay_enabled:
+        if not getattr(self, '_router_replay_enabled', False):
             return lambda: None
 
         self._maybe_apply_router_replay()
diff --git a/src/twinkle/utils/device_mesh.py b/src/twinkle/utils/device_mesh.py
index 7215f7fc..20cda0cf 100644
--- a/src/twinkle/utils/device_mesh.py
+++ b/src/twinkle/utils/device_mesh.py
@@ -538,6 +538,7 @@ class DeviceGroup:
     ranks: Union[List[int], int]
     device_type: str
     gpus_per_worker: int = 1
+    visible_devices: Optional[Union[List[str], List[int], str, int]] = None
     _device_mesh: Dict[str, DeviceMesh] = field(default_factory=dict)