[STABLE ABI] Stable forced_align on cpu

[samanklesaria]

This PR ports the forced_align code to the stable ABI. It uses the temporary Accessor class defined in the stable_accessor PR.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/audio/4022

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 21 New Failures

As of commit 93c6d68 with merge base bdd9c72 ():

NEW FAILURES - The following jobs have failed:

• .github/workflows/bandit.yml (gh)
• .github/workflows/integration-test.yml (gh)
• .github/workflows/lint.yml (gh)
• .github/workflows/pr-labels.yml (gh)
• .github/workflows/unittest-macos-cpu.yml (gh)
• .github/workflows/unittest-windows-cpu.yml (gh)
• .github/workflows/unittest-windows-gpu.yml (gh)
• Build Aarch64 Linux Wheels / pytorch/audio / build-wheel-py3_9-cuda-aarch6412_9-aarch64 (gh)
• Build Aarch64 Linux Wheels / pytorch/audio / upload / upload-wheel-py3_9-cuda-aarch6412_9-aarch64 (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_cu129_aarch64
• Build documentation / build / Build doc (gh)
  RuntimeError: Command docker exec -t e1d8c999f0806918eb9f026700859abf60c78bce3b1026dc93d4f2833448f1ad /exec failed with exit code 1
• Build Linux Wheels / pytorch/audio / build-manywheel-py3_9-cpu (gh)
  Process completed with exit code 1.
• Build Linux Wheels / pytorch/audio / build-manywheel-py3_9-cuda12_6 (gh)
  Process completed with exit code 1.
• Build Linux Wheels / pytorch/audio / build-manywheel-py3_9-cuda12_8 (gh)
  Process completed with exit code 1.
• Build Linux Wheels / pytorch/audio / build-manywheel-py3_9-cuda12_9 (gh)
• Build Linux Wheels / pytorch/audio / build-manywheel-py3_9-xpu (gh)
• Build Linux Wheels / pytorch/audio / upload / upload-manywheel-py3_9-cpu (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_cpu_x86_64
• Build Linux Wheels / pytorch/audio / upload / upload-manywheel-py3_9-cuda12_6 (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_cu126_x86_64
• Build Linux Wheels / pytorch/audio / upload / upload-manywheel-py3_9-cuda12_8 (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_cu128_x86_64
• Build Linux Wheels / pytorch/audio / upload / upload-manywheel-py3_9-cuda12_9 (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_cu129_x86_64
• Build Linux Wheels / pytorch/audio / upload / upload-manywheel-py3_9-xpu (gh)
  Unable to download artifact(s): Artifact not found for name: pytorch_audio__3.9_xpu_x86_64
• Unit-tests on Linux GPU / tests (3.10, 12.6) / linux-job (gh)
  RuntimeError: Command docker exec -t b345276991369c67447648959fdcdeef46eba9404d04aa2faed5552eb5a6d9f4 /exec failed with exit code 1

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[samanklesaria]

Note: this commit now depends on the dtype and is_cpu methods of stable::Tensor in pytorch, which have yet to be merged in. I will keep this PR in draft state until they make their way to nightly.

[samanklesaria]

This PR passes all tests for forced_align on CPU except expected failures. This is because the correct handling of the expected failures requires porting the amax and item ops to the stable ABI. I will port these ops next.

[MahmoudAshraf97]

Hello, is there a chance to get any features or improvements from #3787 merged? on the other side, I managed to remove the torch dependency completely from the FA api, so I guess it's doable, only the input log_probs array needs to be a torch tensor, all other arrays can be vectors or regular arrays, check https://github.com/MahmoudAshraf97/ctc-forced-aligner/blob/main/ctc_forced_aligner/forced_align_impl.cpp for reference

[NicolasHug]

Hi @MahmoudAshraf97 , we're currently trying to migrate these ops to be ABI-stable, which is our current top priority. Unfortunately, we won't have the bandwidth to consider improvements like in #3787 at the moment, sorry. It's possible that we'll be able to re-consider in the future, but the near-term goal is for torchaudio to be stable while minimizing maintenance requirements.

[samanklesaria]

The CI used to run fine. A recent change to the stable ABI now leads to stable::Tensor::scalar_type() being defined twice, which gives all the errors. Perhaps there's a new preferred way of importing the stable API headers?

[pearu]

The CI used to run fine. A recent change to the stable ABI now leads to stable::Tensor::scalar_type() being defined twice, which gives all the errors. Perhaps there's a new preferred way of importing the stable API headers?

The cause of scalar_type() being defined multiple times is in including torch/csrc/stable/tensor.h (read: including torch/csrc/stable/tensor_inl.h that defines scalar_type() as a C++ function) from different files. A workaround (that is used in #4073 and #4079) is to include torch/csrc/stable/tensor.h only from utils.c while all other files include torch/csrc/stable/tensor_struct.h to get the declaration of Tensor.