diff --git a/tutorials/source_en/quick_start/quick_start.md b/tutorials/source_en/quick_start/quick_start.md
index 0ad4925190e0871799c9addc858df112243c53b0..801aa59c62f96c64fc44091af3e335b4d04a390e 100644
--- a/tutorials/source_en/quick_start/quick_start.md
+++ b/tutorials/source_en/quick_start/quick_start.md
@@ -100,6 +100,7 @@ if __name__ == "__main__":
                         help='device where the code will be implemented (default: CPU)')
     args = parser.parse_args()
     context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
+    dataset_sink_mode = not args.device_target == "CPU"
     ...
 ```
 
@@ -338,12 +339,12 @@ from mindspore.train.callback import LossMonitor
 from mindspore.train import Model
 
 ...
-def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb):
+def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, sink_mode):
     """define the training method"""
     print("============== Starting Training ==============")
     #load training dataset
     ds_train = create_dataset(os.path.join(mnist_path, "train"), 32, repeat_size)
-    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=False) # train
+    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=sink_mode) # train
 ...
 
 if __name__ == "__main__":
@@ -353,7 +354,7 @@ if __name__ == "__main__":
     mnist_path = "./MNIST_Data"
     repeat_size = epoch_size
     model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
-    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb)
+    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, dataset_sink_mode)
     ...
 ```
 In the preceding information:    
diff --git a/tutorials/source_zh_cn/quick_start/quick_start.md b/tutorials/source_zh_cn/quick_start/quick_start.md
index f4ffebd70e9d5c714bea5fe571313d09e362a36b..0e83e6752d95ea7dba779fa7251d7ffc8d37e5c2 100644
--- a/tutorials/source_zh_cn/quick_start/quick_start.md
+++ b/tutorials/source_zh_cn/quick_start/quick_start.md
@@ -102,6 +102,7 @@ if __name__ == "__main__":
                         help='device where the code will be implemented (default: CPU)')
     args = parser.parse_args()
     context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
+    dataset_sink_mode = not args.device_target == "CPU"
     ...
 ```
 
@@ -339,12 +340,12 @@ from mindspore.train.callback import LossMonitor
 from mindspore.train import Model
 
 ...
-def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb):
+def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, sink_mode):
     """define the training method"""
     print("============== Starting Training ==============")
     #load training dataset
     ds_train = create_dataset(os.path.join(mnist_path, "train"), 32, repeat_size)
-    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=False)
+    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=sink_mode)
 ...
 
 if __name__ == "__main__":
@@ -354,7 +355,7 @@ if __name__ == "__main__":
     mnist_path = "./MNIST_Data"
     repeat_size = epoch_size
     model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
-    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb)
+    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, dataset_sink_mode)
     ...
 ```
 其中，
diff --git a/tutorials/tutorial_code/lenet.py b/tutorials/tutorial_code/lenet.py
index 84d0a9c54afd3d6cb4f5904b70709644b4a8753a..441f423360c179140f23767970e865e179a10a6c 100644
--- a/tutorials/tutorial_code/lenet.py
+++ b/tutorials/tutorial_code/lenet.py
@@ -169,12 +169,12 @@ class LeNet5(nn.Cell):
         return x
 
 
-def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb):
+def train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, sink_mode):
     """Define the training method."""
     print("============== Starting Training ==============")
     # load training dataset
     ds_train = create_dataset(os.path.join(mnist_path, "train"), 32, repeat_size)
-    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=False)
+    model.train(epoch_size, ds_train, callbacks=[ckpoint_cb, LossMonitor()], dataset_sink_mode=sink_mode)
 
 
 def test_net(args, network, model, mnist_path):
@@ -196,6 +196,7 @@ if __name__ == "__main__":
                         help='device where the code will be implemented (default: CPU)')
     args = parser.parse_args()
     context.set_context(mode=context.GRAPH_MODE, device_target=args.device_target)
+    dataset_sink_mode = not args.device_target == "CPU"
     # download mnist dataset
     download_dataset()
     # learning rate setting
@@ -216,5 +217,5 @@ if __name__ == "__main__":
     # group layers into an object with training and evaluation features
     model = Model(network, net_loss, net_opt, metrics={"Accuracy": Accuracy()})
 
-    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb)
+    train_net(args, model, epoch_size, mnist_path, repeat_size, ckpoint_cb, dataset_sink_mode)
     test_net(args, network, model, mnist_path)