ARM audio support broken

[leondz]

Describe the bug

datasets 4.0 moved to using torchcodec for audio datasets, which doesn't support ARM hardware under Linux and Windows

Steps to reproduce the bug

try:

• Try to install torchcodec under Linux ARM64

or:

• Try to install torchcodec under Windows ARM64

or, on either of those platforms, using datasets >= 4.0:

• dataset = load_dataset("garak-llm/audio_achilles_heel")

Expected behavior

dataset will load as expected

Environment info

ARM environment

Python 3.13+

datasets>=4.0

Using torchaudio

[leondz]

See also NVIDIA/garak#1622 (comment)

[Anakintano]

Hi @leondz , thanks for the detailed bug report — the description and reproduction steps are really clear.

I’d love to take a shot at investigating and potentially fixing this issue. Before I start digging into the codebase, I just wanted to check if there are any additional details, context, or pointers that might help get started (e.g., specific parts of the codebase related to this behavior).

Also, if there are any edge cases or constraints I should be aware of while reproducing the issue, that would be really helpful.

Thanks again for the thorough report!

[lhoestq]

torchcodec works on ARM, see the related discussion on meta-pytorch/torchcodec#569

I see it's missing pure-CPU wheels though

Could you see with the torchcodec community what's missing for you to be able to use torchcodec ?

[leondz]

Thanks, also pursuing there. I want to check: in the interim, is is intended for datasets be broken for audio under aarch64/woa64 when installed with pip, until torchcodec supply appropriate wheels?

[lhoestq]

If the question is: do we plan to implement a temporary workaround in the meantime, the answer is unfortunately no.
This would be a big effort to fix a temporary situation, while using older versions of datasets is actually a simpler workaround.

If older versions of datasets are not working for your case and you really have no other option, I'd be happy to review PRs that offer patches on older versions and I can make a patch release.