array-api compatability

[jorenham]

Typing the array API

Our best bet for proper array-api types is optype, but that doesn't exist yet, see jorenham/optype#25. _{There, I also explain why the official array api "stubs" (which aren't stubs) are completely useless...}

Relevant scipy bits

The Array API is partially supported scipy. The following packages support is fully:

• scipy.cluster
• scipy.constants (trivial)
• scipy.datasets (trivial)
• scipy.fft
  • [i][r|h]fft[2|n]
  • [i]dct[n]
  • [i]dst[n]
  • [r]fftfreq
  • [i]fftshift
• scipy.io
• scipy.ndimage

Additionally, the following scipy.special functions have array api support:

• log_ndtr
• ndtr
• ndtri
• chdtrc
• erf
• erfc
• i0
• i0e
• i1
• i1e
• gammaln
• gammainc
• gammaincc
• logit
• expit
• entr
• rel_entr
• xlogy

as well as the following scipy.stats functions:

• describe
• moment
• skew
• kurtosis
• kstat
• kstatvar
• circmean
• circvar
• circstd
• entropy
• variation
• sem
• ttest_1samp
• pearsonr
• chisquare
• skewtest
• kurtosistest
• normaltest
• jarque_bera
• bartlett
• power_divergence
• monte_carlo_test

How Scipy implements it

Internally, scipy implements the support in scipy._lib._array_api. From the looks of it, this module will come with a bunch of extra utility functions in scipy 1.15 (37174ad), as well as an additional scipy._lib._array_api_no_0d module in 281a8c3

Libraries that support it

Some libraries require array-api-compat for them to support the array-api (2022):

• numpy < 2
• cupy
• pytorch
• dask > 2023.12.0
• ndonnx

the following libraries have native support:

• numpy >= 2
• cupy (but not really)
• jax
• sparse
• ndonnx

Other relevant links:

• https://docs.scipy.org/doc/scipy/dev/api-dev/array_api.html
• https://data-apis.org/array-api/latest/index.html
• https://github.com/data-apis/array-api

[jorenham]

jorenham/optype#170