search

tensorflow / text

Making text a first-class citizen in TensorFlow.
https://www.tensorflow.org/beta/tutorials/tensorflow_text/intro
Apache License 2.0
1.22k stars 334 forks source link

Release aarch64 wheels #1003

Open dwyatte opened 1 year ago

dwyatte commented 1 year ago

With tensorflow 2.10, it looks like TensorFlow is starting to support aarch64. If I recall correctly, tensorflow-text builds for aarch64. Would it be possible to start building those wheels as part of tensorflow-text releases and publishing to PyPI?

broken commented 1 year ago

I expect we should be able to. Let me reach out to the TF infra team to figure how to jump onto this deployment path.

broken commented 1 year ago

After discussing with the core team, that is just a fake wheel that really calls out to the one released by the AWS/ARM/Linaro group. They have an announcement about it here. As such, we are unlikely to add these packages to our releases simply because the build infrastructure isn't there for us.

maparent commented 1 year ago

Apple recently released new versions of tensorflow-macos 2.11 and tensorflow-metal 0.7, with which it was possible to compile tensorflow-text with the usual procedure. I got the resulting wheels for python3.9 and 3.10 to run after building them, but it was very strangely not repeatable, and the wheels are broken. (Apologies for asserting otherwise earlier.)

nicechester commented 2 months ago

I built wheel with python 3.12 and tensorflow 2.16.1 from this source but the procedure wasn't straightforward and it must be done within docker container running on apple silicon. You may try my build if you're facing any issues until there's official binary provided by the team.

pip install https://github.com/nicechester/tensorflow-docker/raw/main/wheels/tensorflow_text-2.16.1-cp312-cp312-linux_aarch64.whl
nouiz commented 1 month ago

Bumping this up. It would be great to have aarch64 wheel for tensorflow-text. This is the only packages that is missing on GraceHopper nodes to install maxtext the same way as on x86.

christinaburge commented 2 weeks ago

Thanks for bumping, we are taking a look and we'll update when we have more info.