tensorstore
tensorstore copied to clipboard
google
Please create a pre-built PyPI wheel for linux arm64
The title says it all...
On https://pypi.org/project/tensorstore/#files there are pre-builts for macosx...arm64 on the one hand, and for manylinux...x86_64 on the other, but none for manylinux...aarch64
See also https://github.com/google/tensorstore/issues/77
We would like to have them, but it is challenging because we currently build on Github-hosted runners which do not include a linux arm64 runner, and building under emulation via qemu is possible but so slow that we decided not to enable that. Cross-compiling would be fast but I could not find a suitable manylinux cross-compilation container with a sufficiently recent GCC version.
Maybe it could be done now by running a linux docker image on the Github-hosted macos arm64 runner.
If gcc 12.2 is sufficiently recent for cross-compilation, it may be worth considering changing the CIBW manylinux containers from manylinux2014 to manylinux_2_28. Additionally, Centos 7, on which manylinux2014 is based, will reach EOL on June 30th 2024.
Yes we were planning to move to manylinux_2_28 anyway for C++20 support. But is there a manylinux_2_28 cross-compilation image with gcc 12?
Yes we were planning to move to manylinux_2_28 anyway for C++20 support. But is there a manylinux_2_28 cross-compilation image with gcc 12?
It looks like cross-compiling aarch64 is still an open issue:
- https://github.com/pypa/cibuildwheel/issues/598
We'd like to see this too for https://github.com/ultralytics/ultralytics
According to our CIs running on Raspberry Pi 5, it takes around 40min+ to build tensorstore. So it would be nice to have prebuilt aarch64 binaries. Thanks :)
https://github.com/ultralytics/ultralytics/actions/runs/9227944458/job/25390932025
FWIW, if you use conda, then conda-forge has tensorstore packages for linux_aarch64.
https://anaconda.org/conda-forge/tensorstore/files
(I understand that may not help if you are not already using conda and don't want to switch.)
