Skip to content

SBSA Support#7484

Open
johnnynunez wants to merge 25 commits into
isl-org:mainfrom
johnnynunez:sbsa_support
Open

SBSA Support#7484
johnnynunez wants to merge 25 commits into
isl-org:mainfrom
johnnynunez:sbsa_support

Conversation

@johnnynunez
Copy link
Copy Markdown
Contributor

@johnnynunez johnnynunez commented May 4, 2026

Summary

  • Upgrade CUDA CI/Docker pins to CUDA 13.2.1 and align Windows CUDA testing with the cu132 PyTorch wheel index.
  • Add CUDA SBSA support via the 5-ml-noble-sbsa Docker config and a new ubuntu-24.04-arm SBSA wheel build job.
  • Split CUDA wheel ccache names for x86_64 and SBSA to avoid cache collisions, and restrict x86 wheel tests to x86_64 artifacts.
  • Refresh CI dependency/action versions, including PyTorch 2.12, TensorFlow 2.21, pip 26.1, and GitHub Actions v6 where used.

Motivation

This prepares Open3D CI for the upcoming PyTorch 2.12 / CUDA 13.2 release and starts producing CUDA SBSA wheels on GitHub’s ARM runners.

Reference: PyTorch 2.12 RC1 announcement

Test plan

  • bash -n docker/docker_build.sh
  • bash -n docker/docker_test.sh
  • Verified the SBSA wheel job tag logic for dev and release modes.
  • Checked lints for the updated workflow and Docker helper files.

@update-docs
Copy link
Copy Markdown

update-docs Bot commented May 4, 2026

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.

@johnnynunez
Copy link
Copy Markdown
Contributor Author

johnnynunez commented May 4, 2026

ready to merge 05/12/2026
it is open now https://download.pytorch.org/whl/cu132

@johnnynunez
Copy link
Copy Markdown
Contributor Author

@ssheorey @benjaminum ready to merge
this enable jetson orin to SBSA, like jetson AGX Thor

The new /Zc:preprocessor flag added for CUDA 13.2 CCCL headers breaks
msgpack-c 3.3.0's MSGPACK_DEFINE_MAP macro expansion under MSVC,
causing all Windows builds to fail. msgpack-cxx 5.0+ fixes this; bump
to 7.0.0 (header-only, drop-in compatible API).
msgpack-cxx 4.0+ includes <boost/predef/other/endian.h> by default for
endianness detection, which Open3D's Linux Docker images do not provide.
The library bundles a Boost-free fallback gated on MSGPACK_NO_BOOST;
propagate that definition through the imported target so all consumers
of <msgpack.hpp> use msgpack/predef/other/endian.h instead.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant