diff --git a/examples/pipeline_tacotron2/README.md b/examples/pipeline_tacotron2/README.md
index a5f30d1d18..5487dc6278 100644
--- a/examples/pipeline_tacotron2/README.md
+++ b/examples/pipeline_tacotron2/README.md
@@ -253,4 +253,4 @@ python inference.py --checkpoint-path ${model_path} \
     --input-text "Hello world!" \
     --text-preprocessor english_characters \
     --output-path "./outputs.wav"
-```
\ No newline at end of file
+```
diff --git a/test/torchaudio_unittest/common_utils/rnnt_utils.py b/test/torchaudio_unittest/common_utils/rnnt_utils.py
index 31887b2beb..94ea300ef7 100644
--- a/test/torchaudio_unittest/common_utils/rnnt_utils.py
+++ b/test/torchaudio_unittest/common_utils/rnnt_utils.py
@@ -5,6 +5,9 @@
 from torchaudio.functional import rnnt_loss
 
 
+CPU_DEVICE = torch.device("cpu")
+
+
 class _NumpyTransducer(torch.autograd.Function):
     @staticmethod
     def forward(
@@ -240,7 +243,7 @@ def get_basic_data(device):
 def get_B1_T10_U3_D4_data(
     random=False,
     dtype=torch.float32,
-    device=torch.device("cpu"),
+    device=CPU_DEVICE,
 ):
     B, T, U, D = 2, 10, 3, 4
 
@@ -263,7 +266,7 @@ def grad_hook(grad):
     return data
 
 
-def get_B1_T2_U3_D5_data(dtype=torch.float32, device=torch.device("cpu")):
+def get_B1_T2_U3_D5_data(dtype=torch.float32, device=CPU_DEVICE):
     logits = torch.tensor(
         [
             0.1,
@@ -360,7 +363,7 @@ def grad_hook(grad):
     return data, ref_costs, ref_gradients
 
 
-def get_B2_T4_U3_D3_data(dtype=torch.float32, device=torch.device("cpu")):
+def get_B2_T4_U3_D3_data(dtype=torch.float32, device=CPU_DEVICE):
     # Test from D21322854
     logits = torch.tensor(
         [
@@ -550,7 +553,7 @@ def get_random_data(
     max_D=40,
     blank=-1,
     dtype=torch.float32,
-    device=torch.device("cpu"),
+    device=CPU_DEVICE,
     seed=None,
 ):
     if seed is not None: