Use torch.accelerator API in Fast Neural Style example

[dvrogozh]

torch.accelerator API allows to abstract some of the accelerator specifics in the user scripts. This commit modifies Fast Neural Style example with the mentioned API.

Things to note:

• Commit modifies example command line by replacing accelerator specific flags (--cuda, --mps, --xpu) with the --accel.
• torch.accelerator API appeared in pytorch 2.6, so commit binds example to this torch version
• Commit introduces USE_ACCEL flag for the test *.sh script. At the moment this flag is just an alias for USE_CUDA. In the future it should replace USE_CUDA once all examples will be converted to use torch.accelerator API.

CC: @malfet, @atalman

[netlify]

✅ Deploy Preview for pytorch-examples-preview canceled.

Name	Link
🔨 Latest commit	afbdd24
🔍 Latest deploy log	https://app.netlify.com/sites/pytorch-examples-preview/deploys/680fd8ceb40ab8000895e151

[msaroufim]

Cool! Although looks like you're not using the right pytorch version quite yet, you can be explicit see how we use requirements.txt in the repo within given folders

[dvrogozh]

I think I might need help with the requirements part of the story. Here is a snapshot of what was executed by ci (https://github.com/pytorch/examples/actions/runs/14539355819/job/40795679639?pr=1327):

2025-04-18T18:32:00.6714511Z Collecting torch
2025-04-18T18:32:00.7041753Z   Downloading torch-2.6.0-cp310-cp310-manylinux1_x86_64.whl.metadata (28 kB)
...
2025-04-18T18:33:28.7956634Z installing requirements
...
2025-04-18T18:33:39.0976123Z Collecting torch>=2 (from torchdata==0.9.0->-r /dev/stdin (line 16))
2025-04-18T18:33:39.1017813Z   Downloading torch-2.5.0-cp310-cp310-manylinux1_x86_64.whl.metadata (28 kB)
...
2025-04-18T18:34:04.9822716Z   Attempting uninstall: torch
2025-04-18T18:34:04.9832302Z     Found existing installation: torch 2.6.0
2025-04-18T18:34:05.3568406Z     Uninstalling torch-2.6.0:
2025-04-18T18:34:05.4241262Z       Successfully uninstalled torch-2.6.0
...
2025-04-18T18:34:37.8267387Z AttributeError: module 'torch' has no attribute 'accelerator'

I think that CI attempted to test the latest nightly, at least this implies that:

examples/.github/workflows/main_python.yml

Line 28 in 8393ceb

pip install --pre torch torchvision -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html

Then, test shell script downgraded already installed nightly torch to version 2.5.0. My guess is due to pinned version of torchvision in some of the requirements.txt files:

examples/dcgan/requirements.txt

Line 2 in 8393ceb

torchvision==0.20.0

Also note:

• Not all examples have requirements.txt file. The example I modified does not have it (check https://github.com/pytorch/examples/tree/main/fast_neural_style)
• CI test scripts actually squash all example requirements.txt files. So, at the moment there is no control which allows to change environment for each sample. Squash happens here:
  

  examples/utils.sh

  Line 26 in 8393ceb

  cat $BASE_DIR/*/requirements.txt | \

[dvrogozh]

I think I might need help with the requirements part of the story.

Filed #1329 to discuss further.

[dvrogozh]

@msaroufim : I've rebased this PR on top of Friday's changes. Note that I suggest to change USE_CUDA to USE_ACCEL in the future once all examples will be converted to the new API. This PR does a step towards that by just adding USE_ACCEL as alias for USE_CUDA.