add an interface for dask.fft

[vtavana]

address #5

[Copilot]

Pull Request Overview

This PR adds a new Dask interface for the mkl_fft package, providing a drop-in replacement for the dask.fft module that leverages Intel MKL's optimized FFT implementations.

• Implements mkl_fft.interfaces.dask_fft module with all FFT functions wrapped using Dask's fft_wrap
• Adds comprehensive test suite for the new Dask interface functionality
• Updates build configuration and documentation to support the new interface

Reviewed Changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
pyproject.toml	Adds dask_interface optional dependency and includes dask in test dependencies
mkl_fft/interfaces/dask_fft.py	New module implementing Dask FFT interface by wrapping numpy_fft functions
mkl_fft/tests/third_party/dask/test_dask_fft.py	Comprehensive test suite for Dask FFT functionality
mkl_fft/tests/test_interfaces.py	Updates interface tests to include Dask FFT
mkl_fft/interfaces/init.py	Conditionally imports dask_fft module when Dask is available
mkl_fft/interfaces/README.md	Adds documentation for the new Dask interface
conda-recipe/meta.yaml	Updates conda recipe to include dask test dependency and import test
conda-recipe-cf/meta.yaml	Updates conda-forge recipe similarly
CHANGELOG.md	Documents the new Dask interface feature
.github/workflows/*.yml	Updates CI workflows to include dask as test dependency

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[jharlow-intel]

some conflicts need to be resolved.

It's okay that the 3.14 internal CI is broken for now, though

[ndgrigorian]

@antonwolfy @jharlow-intel
addressed conflicts, we can get this in now

[ndgrigorian]

changes in the workflows made it more convenient to put this on top of the conda-forge recipe updates, this can go in after that

[ndgrigorian]

@jharlow-intel @antonwolfy ready for review

[jharlow-intel]

Thanks for keeping this going, and sorry it sat so long! The interface itself is close: wrapping mkl_fft's NumPy functions with dask's own fft_wrap keeps the chunking checks and signatures identical to dask.array.fft, and the tests adapted from dask's suite cover chunked-axis errors plus n/s/axes/norm. The main thing is that master has moved a lot since April, so it needs another rebase; a few smaller things inline.

• This needs another rebase before it can merge: master has moved about 330 commits since April, and 7 files now conflict (pyproject.toml, mkl_fft/interfaces/__init__.py, both conda recipes, and the three workflows). When you redo interfaces/__init__.py, it'd be great to follow master's newer pattern there (appending to __all__ and deleting the probe import).

[jharlow-intel]

Could this use pytest.importorskip("dask.array") instead of a bare import? Right now running the installed suite (pytest --pyargs mkl_fft) without dask fails at collection instead of skipping, and test_dask_fftwrap in test_interfaces.py touches mfi.dask_fft unconditionally, so it fails the same way. It might also be worth avoiding the private dask helpers (da._array_expr_enabled, dask.array.numpy_compat), since those can move between dask releases.

[jharlow-intel]

This compares version strings, so "2024.10.0" < "2024.8.2" is True and the norm tests get skipped on dask 2024.10 through 2024.12. packaging.version.Version(dask.__version__) would fix it. And since norm needs dask >= 2024.8.2, should the dask_interface extra pin dask>=2024.8.2 too?

[jharlow-intel]

Since import mkl_fft imports mkl_fft.interfaces eagerly, anyone with dask installed now pulls in dask.array on every import mkl_fft, and a broken dask install would make import mkl_fft itself fail. Would a lazy import (a module-level __getattr__ for dask_fft) work here? I know scipy_fft is eager today, but dask.array is a heavier import.

[jharlow-intel]

Heads-up: this entry ended up in an old released section (next to the Python 3.13 entry, gh-164). Could you move it to the current dev section when you rebase?