Array API adoption Unconference Session

[thomasjpfan]

Scikit-learn and SciPy have been incrementally adding support for the Array API standard. The Array API standard enables libraries to write code in a standard and operate with different array libraries such as NumPy, PyTorch, CuPy, or JAX.

For this unconference session, we will discuss the process of adopting the Array API standard and whether the Array API standard adoption makes sense for your library.

---

When adopting Array API standard, here are some of the common issues that came up:

• How to use integer indexing to set values in an array? x[[0, 3, 4], :] = 3
  • Not part of the standard: https://github.com/data-apis/array-api/issues/177
• You are using NumPy's API that is not a part of the Array API standard?
  • Reimplement the functionality in terms of Array API
    • Need to maintain multiple code paths:
      • scikit-learn has an implementation for nanmin
      • SciPy reimplements [special functions] (https://github.com/scipy/scipy/blob/main/scipy/special/_support_alternative_backends.py) for array libraries that do not implement them and dispatch to libraries that have implementations.
  • Request for an api to be a part of the standard at: https://github.com/data-apis/array-api
• You have compiled code mixed with NumPy-like code?
  • Use the Array API for the NumPy code and dispatch for the compiled code
  • Use dlpack to do zero copy transfer
  • Dispatching is still being developed: NetworkX has a dispatching mechanism
• How to GPU array libraries tests on CI?
  • GitHub actions now have GPU runners

[lucascolley]

I can speak on this topic too as I'm presenting on it at EuroSciPy this Thursday.

[lucascolley]

You have compiled code mixed with NumPy-like code? Use the Array API for the NumPy code and dispatch for the compiled code Use dlpack to do zero copy transfer Dispatching is still being developed: NetworkX has a dispatching mechanism

Some notes on this from discussion at EuroSciPy: https://hackmd.io/@seberg/spatch

[MSanKeys963]

Thanks for proposing the topic, @thomasjpfan!

This is scheduled in Main Room 1. Starts @ 14:40.

Check the schedule here: https://www.nfsummit24.com/schedule.

[lucascolley]

https://github.com/lucascolley/euroscipy24-slides/blob/main/slides.pdf