Machine Learning SIG

Meeting Notes

Next Meeting

• How to increase industry participation
  • AMD/NVIDIA?
  • Google/Meta?
• Should we make this group public?

2023-03-23

• Introductions
• Progress since last meeting
  • PyTorch 2.0: https://github.com/spack/spack/pull/36132
  • TF 2.12: https://github.com/spack/spack/pull/36263
  • tensorboard-data-server Linux aarch64 support: https://github.com/spack/spack/pull/34437
• Pipelines
  • Get PyTorch + ROCm working
  • Add tests for Linux aarch64 and ppc64le

2022-09-13

• Introductions
• Progress since last meeting
  • New JAX package: https://github.com/spack/spack/pull/27849
  • New TF versions: https://github.com/spack/spack/pull/31615, https://github.com/spack/spack/pull/32544
  • TF ROCm support: https://github.com/spack/spack/pull/32248
• Discussion of progress on overall goals
  • https://github.com/orgs/spack/projects/5
• Next steps
  • ML stack in CI: need help from GitLab folks
  • Reach out to Google/Meta to ask about maintainers and build support
  • Reach out to Tamara to ask about distributed/parallel testing in spack test
  • Ask about scientific Python software stack (spider, jupyter, matplotlib, etc.)
  • Concerted effort to fix TF/protobuf linking issues
  • Concerted effort to fix TF build on macOS (x86_64 first, then arm64)

2022-07-27

• Introductions
• Discussion of overall goals of this SIG
  • https://github.com/orgs/spack/projects/5
• How to organize this SIG?
  • Mostly relying on GitHub Projects to track progress
  • Anyone can volunteer/assign themselves to work on a task
  • Communications over GitHub/Slack preferred, email used until we can get everyone on Slack
• How to advertise this SIG?
  • Would love industry partnerships to help maintain build recipes
  • Need some kind of incentive: faster benchmarks?
  • Ideally, every ML package will have both a Spack developer and a ML library developer willing to maintain that build recipe
• Discussion of next steps moving forward
  • Need to decide what exactly we are willing to maintain
    • Which packages?
    • Which platforms?
    • Which versions?
  • Idea: only maintain the packages/platforms/versions we test in CI
  • Scope: only ML packages, not generic data analysis (numpy, pandas) or visualization (matplotlib) tools
• Specific PRs in progress that need reviewing
  • Update TF/Keras: https://github.com/spack/spack/pull/31615
  • ML stack in CI: https://github.com/spack/spack/pull/31592
  • New JAX package: https://github.com/spack/spack/pull/27849