diff --git a/dcgan/README.md b/dcgan/README.md
index 3f7bdef6b1..b921bb7e22 100644
--- a/dcgan/README.md
+++ b/dcgan/README.md
@@ -24,7 +24,7 @@ usage: main.py [-h] --dataset DATASET --dataroot DATAROOT [--workers WORKERS]
                [--batchSize BATCHSIZE] [--imageSize IMAGESIZE] [--nz NZ]
                [--ngf NGF] [--ndf NDF] [--niter NITER] [--lr LR]
                [--beta1 BETA1] [--cuda] [--ngpu NGPU] [--netG NETG]
-               [--netD NETD] [--mps]
+               [--netD NETD] [--mps] [--device DEVICE]
 
 optional arguments:
   -h, --help            show this help message and exit
@@ -41,6 +41,7 @@ optional arguments:
   --beta1 BETA1         beta1 for adam. default=0.5
   --cuda                enables cuda
   --mps                 enables macOS GPU
+  --device              backend device
   --ngpu NGPU           number of GPUs to use
   --netG NETG           path to netG (to continue training)
   --netD NETD           path to netD (to continue training)
diff --git a/dcgan/main.py b/dcgan/main.py
index 2f45b2dbd2..9416893856 100644
--- a/dcgan/main.py
+++ b/dcgan/main.py
@@ -34,6 +34,7 @@
 parser.add_argument('--manualSeed', type=int, help='manual seed')
 parser.add_argument('--classes', default='bedroom', help='comma separated list of classes for the lsun data set')
 parser.add_argument('--mps', action='store_true', default=False, help='enables macOS GPU training')
+parser.add_argument('--device', type=str, default='cpu', help='backend device')
 
 opt = parser.parse_args()
 print(opt)
@@ -112,7 +113,7 @@
 elif use_mps:
     device = torch.device("mps")
 else:
-    device = torch.device("cpu")
+    device = torch.device(opt.device)
 
 ngpu = int(opt.ngpu)
 nz = int(opt.nz)
diff --git a/fast_neural_style/README.md b/fast_neural_style/README.md
index 8057847214..674cf7e7f0 100644
--- a/fast_neural_style/README.md
+++ b/fast_neural_style/README.md
@@ -28,6 +28,7 @@ python neural_style/neural_style.py eval --content-image </path/to/content/image
 - `--content-scale`: factor for scaling down the content image if memory is an issue (eg: value of 2 will halve the height and width of content-image)
 - `--cuda`: set it to 1 for running on GPU, 0 for CPU.
 - `--mps`: set it to 1 for running on macOS GPU
+- `--device DEVICE`: backend device to run on, 'cpu' by default.
 
 Train model
 
@@ -42,6 +43,7 @@ There are several command line arguments, the important ones are listed below
 - `--save-model-dir`: path to folder where trained model will be saved.
 - `--cuda`: set it to 1 for running on GPU, 0 for CPU.
 - `--mps`: set it to 1 for running on macOS GPU
+- `--device DEVICE`: backend device to run on, 'cpu' by default.
 
 Refer to `neural_style/neural_style.py` for other command line arguments. For training new models you might have to tune the values of `--content-weight` and `--style-weight`. The mosaic style model shown above was trained with `--content-weight 1e5` and `--style-weight 1e10`. The remaining 3 models were also trained with similar order of weight parameters with slight variation in the `--style-weight` (`5e10` or `1e11`).
 
diff --git a/fast_neural_style/neural_style/neural_style.py b/fast_neural_style/neural_style/neural_style.py
index 91bf642d82..b1cd2c7687 100644
--- a/fast_neural_style/neural_style/neural_style.py
+++ b/fast_neural_style/neural_style/neural_style.py
@@ -34,7 +34,7 @@ def train(args):
     elif args.mps:
         device = torch.device("mps")
     else:
-        device = torch.device("cpu")
+        device = torch.device(args.device)
 
     np.random.seed(args.seed)
     torch.manual_seed(args.seed)
@@ -125,7 +125,7 @@ def train(args):
 
 
 def stylize(args):
-    device = torch.device("cuda" if args.cuda else "cpu")
+    device = torch.device("cuda" if args.cuda else args.device)
 
     content_image = utils.load_image(args.content_image, scale=args.content_scale)
     content_transform = transforms.Compose([
@@ -205,8 +205,10 @@ def main():
                                   help="size of training images, default is 256 X 256")
     train_arg_parser.add_argument("--style-size", type=int, default=None,
                                   help="size of style-image, default is the original size of style image")
-    train_arg_parser.add_argument("--cuda", type=int, required=True,
+    train_arg_parser.add_argument("--cuda", type=int, required=True, default=0,
                                   help="set it to 1 for running on GPU, 0 for CPU")
+    train_arg_parser.add_argument('--device', type=str, default='cpu',
+                                  help='backend device')
     train_arg_parser.add_argument("--seed", type=int, default=42,
                                   help="random seed for training")
     train_arg_parser.add_argument("--content-weight", type=float, default=1e5,
@@ -234,6 +236,7 @@ def main():
     eval_arg_parser.add_argument("--export_onnx", type=str,
                                  help="export ONNX model to a given file")
     eval_arg_parser.add_argument('--mps', action='store_true', default=False, help='enable macOS GPU training')
+    eval_arg_parser.add_argument('--device', type=str, default='cpu', help='backend device')
 
     args = main_arg_parser.parse_args()
 
diff --git a/gat/README.md b/gat/README.md
index 7bb71bc17b..f4d6848331 100644
--- a/gat/README.md
+++ b/gat/README.md
@@ -69,7 +69,7 @@ python main.py --epochs 300 --lr 0.005 --l2 5e-4 --dropout-p 0.6 --num-heads 8 -
 In more detail, the `main.py` script recieves following arguments:
 ```
 usage: main.py [-h] [--epochs EPOCHS] [--lr LR] [--l2 L2] [--dropout-p DROPOUT_P] [--hidden-dim HIDDEN_DIM] [--num-heads NUM_HEADS] [--concat-heads] [--val-every VAL_EVERY]
-               [--no-cuda] [--no-mps] [--dry-run] [--seed S]
+               [--no-cuda] [--no-mps] [--dry-run] [--seed S] [--device DEVICE]
 
 PyTorch Graph Attention Network
 
@@ -89,6 +89,7 @@ options:
                         epochs to wait for print training and validation evaluation (default: 20)
   --no-cuda             disables CUDA training
   --no-mps              disables macOS GPU training
+  --device DEVICE       backend device
   --dry-run             quickly check a single pass
   --seed S              random seed (default: 13)
 ```
diff --git a/gat/main.py b/gat/main.py
index 9c143af8ec..d70e380c94 100644
--- a/gat/main.py
+++ b/gat/main.py
@@ -311,6 +311,8 @@ def test(model, criterion, input, target, mask):
                         help='disables CUDA training')
     parser.add_argument('--no-mps', action='store_true', default=False,
                         help='disables macOS GPU training')
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--dry-run', action='store_true', default=False,
                         help='quickly check a single pass')
     parser.add_argument('--seed', type=int, default=13, metavar='S',
@@ -327,7 +329,7 @@ def test(model, criterion, input, target, mask):
     elif use_mps:
         device = torch.device('mps')
     else:
-        device = torch.device('cpu')
+        device = torch.device(args.device)
     print(f'Using {device} device')
 
     # Load the dataset
diff --git a/gcn/main.py b/gcn/main.py
index 5c8362b576..049129f64b 100644
--- a/gcn/main.py
+++ b/gcn/main.py
@@ -220,6 +220,8 @@ def test(model, criterion, input, target, mask):
                         help='disables CUDA training')
     parser.add_argument('--no-mps', action='store_true', default=False,
                         help='disables macOS GPU training')
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--dry-run', action='store_true', default=False,
                         help='quickly check a single pass')
     parser.add_argument('--seed', type=int, default=42, metavar='S',
@@ -236,7 +238,7 @@ def test(model, criterion, input, target, mask):
     elif use_mps:
         device = torch.device('mps')
     else:
-        device = torch.device('cpu')
+        device = torch.device(args.device)
     print(f'Using {device} device')
 
     cora_url = 'https://linqs-data.soe.ucsc.edu/public/lbc/cora.tgz'
diff --git a/language_translation/main.py b/language_translation/main.py
index 2b4fbb94c3..3e82748556 100644
--- a/language_translation/main.py
+++ b/language_translation/main.py
@@ -272,9 +272,9 @@ def main(opts):
                         help="Default learning rate")
     parser.add_argument("--batch", type=int, default=128,
                         help="Batch size")
-    parser.add_argument("--backend", type=str, default="cpu",
-                        help="Batch size")
-    
+    parser.add_argument("--device", type=str, default="cpu",
+                        help="backend device")
+
     # Transformer settings
     parser.add_argument("--attn_heads", type=int, default=8,
                         help="Number of attention heads")
@@ -298,7 +298,7 @@ def main(opts):
 
     args = parser.parse_args()
 
-    DEVICE = torch.device("cuda" if args.backend == "gpu" and torch.cuda.is_available() else "cpu")
+    DEVICE = torch.device("cuda" if args.device == "gpu" and torch.cuda.is_available() else args.device)
 
     if args.inference:
         inference(args)
diff --git a/legacy/snli/README.md b/legacy/snli/README.md
index 2da5104975..a7d2927c10 100644
--- a/legacy/snli/README.md
+++ b/legacy/snli/README.md
@@ -25,7 +25,7 @@ spacy
 Start the training process with:
 
 ```bash
-python train.py --lower --word-vectors [PATH_TO_WORD_VECTORS] --vector-cache [PATH_TO_VECTOR_CACHE] --epochs [NUMBER_OF_EPOCHS] --batch-size [BATCH_SIZE] --save-path [PATH_TO_SAVE_MODEL] --gpu [GPU_NUMBER]
+python train.py --lower --word-vectors [PATH_TO_WORD_VECTORS] --vector-cache [PATH_TO_VECTOR_CACHE] --epochs [NUMBER_OF_EPOCHS] --batch-size [BATCH_SIZE] --save-path [PATH_TO_SAVE_MODEL] --gpu [GPU_NUMBER] --device [BACKEND_DEVICE]
 ```
 
 ## 🏋️‍♀️ Training
diff --git a/legacy/snli/train.py b/legacy/snli/train.py
index aa70aef45a..f664ffdfbd 100644
--- a/legacy/snli/train.py
+++ b/legacy/snli/train.py
@@ -20,7 +20,7 @@
 elif torch.backends.mps.is_available():
     device = torch.device('mps')
 else:
-    device = torch.device('cpu')
+    device = torch.device(args.device)
 
 inputs = data.Field(lower=args.lower, tokenize='spacy')
 answers = data.Field(sequential=False)
diff --git a/legacy/snli/util.py b/legacy/snli/util.py
index 1bc8e0b2cc..4e2b488767 100644
--- a/legacy/snli/util.py
+++ b/legacy/snli/util.py
@@ -20,6 +20,8 @@ def makedirs(name):
 
 def get_args():
     parser = ArgumentParser(description='PyTorch/torchtext SNLI example')
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--epochs', type=int, default=50,
                         help='the number of total epochs to run.')
     parser.add_argument('--batch_size', type=int, default=128,
diff --git a/mnist/main.py b/mnist/main.py
index 184dc4744f..d00cbb67c0 100644
--- a/mnist/main.py
+++ b/mnist/main.py
@@ -86,6 +86,8 @@ def main():
                         help='disables CUDA training')
     parser.add_argument('--no-mps', action='store_true', default=False,
                         help='disables macOS GPU training')
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--dry-run', action='store_true', default=False,
                         help='quickly check a single pass')
     parser.add_argument('--seed', type=int, default=1, metavar='S',
@@ -105,7 +107,7 @@ def main():
     elif use_mps:
         device = torch.device("mps")
     else:
-        device = torch.device("cpu")
+        device = torch.device(args.device)
 
     train_kwargs = {'batch_size': args.batch_size}
     test_kwargs = {'batch_size': args.test_batch_size}
diff --git a/mnist_forward_forward/README.md b/mnist_forward_forward/README.md
index f6ae12e56d..6a8f70aace 100644
--- a/mnist_forward_forward/README.md
+++ b/mnist_forward_forward/README.md
@@ -18,6 +18,7 @@ optional arguments:
   --lr LR               learning rate (default: 0.03)
   --no_cuda             disables CUDA training
   --no_mps              disables MPS training
+  --device DEVICE       backend device
   --seed SEED           random seed (default: 1)
   --save_model          For saving the current Model
   --train_size TRAIN_SIZE
diff --git a/mnist_forward_forward/main.py b/mnist_forward_forward/main.py
index f137dee48a..3702b213ea 100644
--- a/mnist_forward_forward/main.py
+++ b/mnist_forward_forward/main.py
@@ -108,6 +108,9 @@ def train(self, x_pos, x_neg):
     parser.add_argument(
         "--no_mps", action="store_true", default=False, help="disables MPS training"
     )
+    parser.add_argument(
+        '--device', type=str, default='cpu', help='backend device'
+    )
     parser.add_argument(
         "--seed", type=int, default=1, metavar="S", help="random seed (default: 1)"
     )
@@ -145,7 +148,7 @@ def train(self, x_pos, x_neg):
     elif use_mps:
         device = torch.device("mps")
     else:
-        device = torch.device("cpu")
+        device = torch.device(args.device)
 
     train_kwargs = {"batch_size": args.train_size}
     test_kwargs = {"batch_size": args.test_size}
diff --git a/mnist_hogwild/README.md b/mnist_hogwild/README.md
index 5f12161d53..361ff43415 100644
--- a/mnist_hogwild/README.md
+++ b/mnist_hogwild/README.md
@@ -21,6 +21,7 @@ optional arguments:
   --log_interval        how many batches to wait before logging training status
   --num_process         how many training processes to use (default: 2)
   --cuda                enables CUDA training
+  --device DEVICE       backend device
   --dry-run             quickly check a single pass
   --save-model          For Saving the current Model
 ```
diff --git a/mnist_hogwild/main.py b/mnist_hogwild/main.py
index 6fa449233d..6af0c4da71 100644
--- a/mnist_hogwild/main.py
+++ b/mnist_hogwild/main.py
@@ -31,6 +31,8 @@
                     help='enables CUDA training')
 parser.add_argument('--mps', action='store_true', default=False,
                     help='enables macOS GPU training')
+parser.add_argument('--device', type=str, default='cpu',
+                    help='backend device')
 parser.add_argument('--save_model', action='store_true', default=False,
                     help='save the trained model to state_dict')
 parser.add_argument('--dry-run', action='store_true', default=False,
@@ -65,7 +67,7 @@ def forward(self, x):
     elif use_mps:
         device = torch.device("mps")
     else:
-        device = torch.device("cpu")
+        device = torch.device(args.device)
 
     transform=transforms.Compose([
         transforms.ToTensor(),
diff --git a/mnist_rnn/main.py b/mnist_rnn/main.py
index 2fa64c00d6..c77191ce8c 100644
--- a/mnist_rnn/main.py
+++ b/mnist_rnn/main.py
@@ -95,6 +95,8 @@ def main():
                         help='enables CUDA training')
     parser.add_argument('--mps', action="store_true", default=False,
                         help="enables MPS training")
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--dry-run', action='store_true', default=False,
                         help='quickly check a single pass')
     parser.add_argument('--seed', type=int, default=1, metavar='S',
@@ -110,7 +112,7 @@ def main():
     elif args.mps and not args.cuda:
         device = "mps"
     else:
-        device = "cpu"
+        device = args.device
 
     device = torch.device(device)
 
diff --git a/run_python_examples.sh b/run_python_examples.sh
index 0e06e4cfc0..344846e95e 100755
--- a/run_python_examples.sh
+++ b/run_python_examples.sh
@@ -13,6 +13,15 @@
 BASE_DIR="$(pwd)/$(dirname $0)"
 source $BASE_DIR/utils.sh
 
+# Run on a specific backend device with 'export BACKEND_DEVICE=cpu'. It will
+# be set to 'cpu' by default and has lower priority than '--cuda' and '--mps'.
+# See https://github.com/pytorch/examples/pull/1288 for more information.
+if [ -n "${BACKEND_DEVICE}" ]; then
+  DEVICE_FLAG="--device ${BACKEND_DEVICE}"
+else
+  DEVICE_FLAG=""
+fi
+
 USE_CUDA=$(python -c "import torchvision, torch; print(torch.cuda.is_available())")
 case $USE_CUDA in
   "True")
@@ -32,7 +41,7 @@ esac
 
 function dcgan() {
   start
-  python main.py --dataset fake $CUDA_FLAG --mps --dry-run || error "dcgan failed"
+  python main.py --dataset fake $CUDA_FLAG --mps $DEVICE_FLAG --dry-run || error "dcgan failed"
 }
 
 function fast_neural_style() {
@@ -44,7 +53,7 @@ function fast_neural_style() {
   test -d "saved_models" || { error "saved models not found"; return; }
 
   echo "running fast neural style model"
-  python neural_style/neural_style.py eval --content-image images/content-images/amber.jpg --model saved_models/candy.pth --output-image images/output-images/amber-candy.jpg --cuda $CUDA --mps || error "neural_style.py failed"
+  python neural_style/neural_style.py eval --content-image images/content-images/amber.jpg --model saved_models/candy.pth --output-image images/output-images/amber-candy.jpg --cuda $CUDA --mps $DEVICE_FLAG || error "neural_style.py failed"
 }
 
 function imagenet() {
@@ -63,36 +72,36 @@ function language_translation() {
   start
   python -m spacy download en || error "couldn't download en package from spacy"
   python -m spacy download de || error "couldn't download de package from spacy"
-  python main.py -e 1 --enc_layers 1 --dec_layers 1 --backend cpu --logging_dir output/ --dry_run || error "language translation example failed"
+  python main.py -e 1 --enc_layers 1 --dec_layers 1 $DEVICE_FLAG --logging_dir output/ --dry_run || error "language translation example failed"
 }
 
 function mnist() {
   start
-  python main.py --epochs 1 --dry-run || error "mnist example failed"
+  python main.py --epochs 1 --dry-run $DEVICE_FLAG || error "mnist example failed"
 }
 function mnist_forward_forward() {
   start
-  python main.py --epochs 1 --no_mps --no_cuda || error "mnist forward forward failed"
+  python main.py --epochs 1 --no_mps --no_cuda $DEVICE_FLAG || error "mnist forward forward failed"
 
 }
 function mnist_hogwild() {
   start
-  python main.py --epochs 1 --dry-run $CUDA_FLAG || error "mnist hogwild failed"
+  python main.py --epochs 1 --dry-run $CUDA_FLAG $DEVICE_FLAG || error "mnist hogwild failed"
 }
 
 function mnist_rnn() {
   start
-  python main.py --epochs 1 --dry-run || error "mnist rnn example failed"
+  python main.py --epochs 1 --dry-run $DEVICE_FLAG || error "mnist rnn example failed"
 }
 
 function regression() {
   start
-  python main.py --epochs 1 $CUDA_FLAG || error "regression failed"
+  python main.py --epochs 1 $CUDA_FLAG $DEVICE_FLAG || error "regression failed"
 }
 
 function siamese_network() {
   start
-  python main.py --epochs 1 --dry-run || error "siamese network example failed"
+  python main.py --epochs 1 --dry-run $DEVICE_FLAG || error "siamese network example failed"
 }
 
 function reinforcement_learning() {
@@ -123,7 +132,7 @@ function fx() {
 
 function super_resolution() {
   start
-  python main.py --upscale_factor 3 --batchSize 4 --testBatchSize 100 --nEpochs 1 --lr 0.001 --mps || error "super resolution failed"
+  python main.py --upscale_factor 3 --batchSize 4 --testBatchSize 100 --nEpochs 1 --lr 0.001 --mps $DEVICE_FLAG || error "super resolution failed"
 }
 
 function time_sequence_prediction() {
@@ -134,7 +143,7 @@ function time_sequence_prediction() {
 
 function vae() {
   start
-  python main.py --epochs 1 || error "vae failed"
+  python main.py --epochs 1 $DEVICE_FLAG || error "vae failed"
 }
 
 function vision_transformer() {
@@ -144,17 +153,17 @@ function vision_transformer() {
 
 function word_language_model() {
   start
-  python main.py --epochs 1 --dry-run $CUDA_FLAG --mps || error "word_language_model failed"
+  python main.py --epochs 1 --dry-run $CUDA_FLAG --mps $DEVICE_FLAG || error "word_language_model failed"
 }
 
 function gcn() {
   start
-  python main.py --epochs 1 --dry-run || error "graph convolutional network failed"
+  python main.py --epochs 1 --dry-run $DEVICE_FLAG || error "graph convolutional network failed"
 }
 
 function gat() {
   start
-  python main.py --epochs 1 --dry-run || error "graph attention network failed"
+  python main.py --epochs 1 --dry-run $DEVICE_FLAG || error "graph attention network failed"
 }
 
 function clean() {
diff --git a/siamese_network/main.py b/siamese_network/main.py
index 8f420a9b01..1d72653a60 100644
--- a/siamese_network/main.py
+++ b/siamese_network/main.py
@@ -251,6 +251,8 @@ def main():
                         help='disables CUDA training')
     parser.add_argument('--no-mps', action='store_true', default=False,
                         help='disables macOS GPU training')
+    parser.add_argument('--device', type=str, default='cpu',
+                        help='backend device')
     parser.add_argument('--dry-run', action='store_true', default=False,
                         help='quickly check a single pass')
     parser.add_argument('--seed', type=int, default=1, metavar='S',
@@ -271,7 +273,7 @@ def main():
     elif use_mps:
         device = torch.device("mps")
     else:
-        device = torch.device("cpu")
+        device = torch.device(args.device)
 
     train_kwargs = {'batch_size': args.batch_size}
     test_kwargs = {'batch_size': args.test_batch_size}
diff --git a/super_resolution/README.md b/super_resolution/README.md
index 6b5fe831d9..883a9b7d12 100644
--- a/super_resolution/README.md
+++ b/super_resolution/README.md
@@ -5,7 +5,7 @@ This example illustrates how to use the efficient sub-pixel convolution layer de
 ```
 usage: main.py [-h] --upscale_factor UPSCALE_FACTOR [--batchSize BATCHSIZE]
                [--testBatchSize TESTBATCHSIZE] [--nEpochs NEPOCHS] [--lr LR]
-               [--cuda] [--threads THREADS] [--seed SEED]
+               [--cuda] [--device DEVICE] [--threads THREADS] [--seed SEED]
 
 PyTorch Super Res Example
 
@@ -18,6 +18,7 @@ optional arguments:
   --lr                  Learning Rate. Default=0.01
   --cuda                use cuda
   --mps                 enable GPU on macOS
+  --device DEVICE       backend device
   --threads             number of threads for data loader to use Default=4
   --seed                random seed to use. Default=123
 ```
diff --git a/super_resolution/main.py b/super_resolution/main.py
index 8c5519582a..8b6850706a 100644
--- a/super_resolution/main.py
+++ b/super_resolution/main.py
@@ -18,6 +18,7 @@
 parser.add_argument('--lr', type=float, default=0.01, help='Learning Rate. Default=0.01')
 parser.add_argument('--cuda', action='store_true', help='use cuda?')
 parser.add_argument('--mps', action='store_true', default=False, help='enables macOS GPU training')
+parser.add_argument('--device', type=str, default='cpu', help='backend device')
 parser.add_argument('--threads', type=int, default=4, help='number of threads for data loader to use')
 parser.add_argument('--seed', type=int, default=123, help='random seed to use. Default=123')
 opt = parser.parse_args()
@@ -37,7 +38,7 @@
 elif use_mps:
     device = torch.device("mps")
 else:
-    device = torch.device("cpu")
+    device = torch.device(opt.device)
 
 print('===> Loading datasets')
 train_set = get_training_set(opt.upscale_factor)
diff --git a/super_resolution/super_resolve.py b/super_resolution/super_resolve.py
index 750d635312..2db3d5297d 100644
--- a/super_resolution/super_resolve.py
+++ b/super_resolution/super_resolve.py
@@ -12,6 +12,7 @@
 parser.add_argument('--model', type=str, required=True, help='model file to use')
 parser.add_argument('--output_filename', type=str, help='where to save the output image')
 parser.add_argument('--cuda', action='store_true', help='use cuda')
+parser.add_argument('--device', type=str, default='cpu', help='backend device')
 opt = parser.parse_args()
 
 print(opt)
@@ -25,6 +26,9 @@
 if opt.cuda:
     model = model.cuda()
     input = input.cuda()
+elif opt.device:
+    model = model.to(opt.device)
+    input = input.to(opt.device)
 
 out = model(input)
 out = out.cpu()
diff --git a/vae/README.md b/vae/README.md
index cda6a33672..393776f219 100644
--- a/vae/README.md
+++ b/vae/README.md
@@ -12,10 +12,11 @@ The main.py script accepts the following arguments:
 
 ```bash
 optional arguments:
-  --batch-size		input batch size for training (default: 128)
-  --epochs		number of epochs to train (default: 10)
-  --no-cuda		enables CUDA training
-  --mps         enables GPU on macOS
-  --seed		random seed (default: 1)
-  --log-interval	how many batches to wait before logging training status
-```
\ No newline at end of file
+  --batch-size N        input batch size for training (default: 128)
+  --epochs EPOCHS       number of epochs to train (default: 10)
+  --no-cuda             disables CUDA training
+  --no-mps              disables macOS GPU training
+  --device DEVICE       backend name
+  --seed SEED           random seed (default: 1)
+  --log-interval N      how many batches to wait before logging training status
+```
diff --git a/vae/main.py b/vae/main.py
index d69833fbe0..a49efe4fdd 100644
--- a/vae/main.py
+++ b/vae/main.py
@@ -17,6 +17,8 @@
                     help='disables CUDA training')
 parser.add_argument('--no-mps', action='store_true', default=False,
                         help='disables macOS GPU training')
+parser.add_argument('--device', type=str, default='cpu',
+                    help='backend device')
 parser.add_argument('--seed', type=int, default=1, metavar='S',
                     help='random seed (default: 1)')
 parser.add_argument('--log-interval', type=int, default=10, metavar='N',
@@ -32,7 +34,7 @@
 elif use_mps:
     device = torch.device("mps")
 else:
-    device = torch.device("cpu")
+    device = torch.device(args.device)
 
 kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
 train_loader = torch.utils.data.DataLoader(
diff --git a/word_language_model/README.md b/word_language_model/README.md
index 254b726585..ca217399c8 100644
--- a/word_language_model/README.md
+++ b/word_language_model/README.md
@@ -37,6 +37,7 @@ optional arguments:
   --seed SEED           random seed
   --cuda                use CUDA
   --mps                 enable GPU on macOS
+  --device DEVICE       backend device
   --log-interval N      report interval
   --save SAVE           path to save the final model
   --onnx-export ONNX_EXPORT
@@ -54,3 +55,9 @@ python main.py --cuda --emsize 650 --nhid 650 --dropout 0.5 --epochs 40 --tied
 python main.py --cuda --emsize 1500 --nhid 1500 --dropout 0.65 --epochs 40
 python main.py --cuda --emsize 1500 --nhid 1500 --dropout 0.65 --epochs 40 --tied
 ```
+
+You can also use non-cuda devices like this:
+
+```bash
+python main.py --device npu --emsize 650 --nhid 650 --dropout 0.5 --epochs 40
+```
diff --git a/word_language_model/generate.py b/word_language_model/generate.py
index 13bd8abfcd..c201e3ff21 100644
--- a/word_language_model/generate.py
+++ b/word_language_model/generate.py
@@ -23,6 +23,8 @@
                     help='random seed')
 parser.add_argument('--cuda', action='store_true',
                     help='use CUDA')
+parser.add_argument('--device', type=str, default='cpu',
+                    help='backend device')
 parser.add_argument('--mps', action='store_true', default=False,
                         help='enables macOS GPU training')
 parser.add_argument('--temperature', type=float, default=1.0,
@@ -46,7 +48,7 @@
 elif use_mps:
     device = torch.device("mps")
 else:
-    device = torch.device("cpu")
+    device = torch.device(args.device)
 
 if args.temperature < 1e-3:
     parser.error("--temperature has to be greater or equal 1e-3.")
diff --git a/word_language_model/main.py b/word_language_model/main.py
index 23bda03e73..71c4446003 100644
--- a/word_language_model/main.py
+++ b/word_language_model/main.py
@@ -40,7 +40,9 @@
 parser.add_argument('--cuda', action='store_true', default=False,
                     help='use CUDA')
 parser.add_argument('--mps', action='store_true', default=False,
-                        help='enables macOS GPU training')
+                    help='enables macOS GPU training')
+parser.add_argument('--device', type=str, default='cpu',
+                    help='backend device')
 parser.add_argument('--log-interval', type=int, default=200, metavar='N',
                     help='report interval')
 parser.add_argument('--save', type=str, default='model.pt',
@@ -68,7 +70,7 @@
 elif use_mps:
     device = torch.device("mps")
 else:
-    device = torch.device("cpu")
+    device = torch.device(args.device)
 
 ###############################################################################
 # Load data