diff --git a/ppcls/arch/backbone/legendary_models/swin_transformer.py b/ppcls/arch/backbone/legendary_models/swin_transformer.py
index 412d76ba8..50d781c96 100644
--- a/ppcls/arch/backbone/legendary_models/swin_transformer.py
+++ b/ppcls/arch/backbone/legendary_models/swin_transformer.py
@@ -62,7 +62,6 @@ class RollWithIndexSelect(paddle.autograd.PyLayer):
         grad_input = grad.reshape([N, H * W, C]).index_select(index_bp, 1).reshape([N, H, W, C])
         return grad_input, None, None
 
-roll_with_index_select = RollWithIndexSelect.apply
 
 def get_roll_index(H, W, shifts, place):
     # following tensors will be created on cpu place with npu custom device
@@ -74,11 +73,26 @@ def get_roll_index(H, W, shifts, place):
     index_bp = paddle.to_tensor(index_fp, dtype='int64', place=place)
     return [index_fp, index_bp]
 
-class NpuRollWithIndexSelect():
+def singleton(class_):
+    instances = {}
+    def getinstance(*args, **kwargs):
+        if class_ not in instances:
+            instances[class_] = class_(*args, **kwargs)
+        return instances[class_]
+    return getinstance
+
+@singleton
+class RollWrapperSingleton():
     def __init__(self):
         self.index_dict = {}
+        self.roll_with_index_select = RollWithIndexSelect.apply
+        if 'npu' in paddle.device.get_all_custom_device_type():
+            self.enable = True
 
     def __call__(self, x, shifts, axis):
+        if not self.enable:
+            return padlde.roll(x, shifts, axis)
+
         assert x.dim() == 4
         assert len(shifts) == 2
         assert len(axis) == 2
@@ -87,16 +101,8 @@ class NpuRollWithIndexSelect():
         if key not in self.index_dict:
             self.index_dict[key] = get_roll_index(H, W, shifts, x.place)
         index_fp, index_bp = self.index_dict[key]
-        return roll_with_index_select(x, index_fp, index_bp)
+        return self.roll_with_index_select(x, index_fp, index_bp)
 
-roll = None
-
-def _lazy_init_roll(x):
-    global roll
-    if 'npu' in paddle.device.get_all_custom_device_type() and hasattr(x, '_place_str') and 'npu' in x._place_str:
-        roll = NpuRollWithIndexSelect()
-    else:
-        roll = paddle.roll
 
 class Mlp(nn.Layer):
     def __init__(self,
@@ -409,11 +415,9 @@ class SwinTransformerBlock(nn.Layer):
         x = self.norm1(x)
         x = x.reshape([B, H, W, C])
 
+        roll = RollWrapperSingleton()
         # cyclic shift
         if self.shift_size > 0:
-            if roll is None:
-                _lazy_init_roll(x)
-
             shifted_x = roll(
                 x, shifts=(-self.shift_size, -self.shift_size), axis=(1, 2))
         else:
@@ -438,9 +442,6 @@ class SwinTransformerBlock(nn.Layer):
 
         # reverse cyclic shift
         if self.shift_size > 0:
-            if roll is None:
-                _lazy_init_roll(shifted_x)
-
             x = roll(
                 shifted_x,
                 shifts=(self.shift_size, self.shift_size),