This is my first venture into array-api-extra land.

welcome :)

Given that this was mostly a "copy&paste" job I thought it would be a good task to learn more about using "agents", so I asked Cursor to do the work, etc. From what I can tell reviewing the diff it looks sensible. But the caveat is that I've never thought about how to write quantile myself. So you might consider this "the blind leading the blind". There are a few differences to the scipy implementation, mostly more "make this an array" and dealing with squashing/maintaining dimension handling. I can't explain to you why these differences make sense :-/ - both implementations seem sensible when looked at individually. Maybe someone here can explain how that fits with the fact that the diff isn't zero :D

I would rather merge any changes to the implementation into SciPy first, before diverging into uncharted territory with LLM magic here.

A few of the pre-commit checks related to numpydoc conventions don't understand the idea of only having a full docstring in _delegation.py and not also in _funcs.py - I turned off the checks when there were no other errors left. I'm not sure I get the point of some of the ruff checks that failed, hence the noqa comments on the raise statements (can change them to assign the message to a variable first)

Sounds fine, existing noqa and numpydoc: ignore comments should be a good guide to what we have typically been doing so far.

I appeased the linter for you

I am breaking out some of the lint changes into gh-344

diff all tidied up

With the changes to the liniting/pre-commit setup: is https://github.com/evilmartians/lefthook the lefthook you are now using? I think the docs/pyproject.toml need tidying up because it still refers to pre-commit but I think that is completely gone now.

And/or: it would be nice to have instructions for how to run the required tools to make the CI checks happy that doesn't involve tools calling tools calling tools. I'm not sure I care enough about pixi (or you know one of the five other tools-that-call-tools that other projects use) to learn enough about it to figure out if I can safely install it in a conda env and it won't touch anything outside of that env - or if it needs installing "globally" or in parallel to conda env etc. This isn't specific to pixi, but more because five different repos use six different of these tools :-/

I think the docs/pyproject.toml need tidying up because it still refers to pre-commit but I think that is completely gone now.

Please point out what seems to need tidying up! Bear in mind that the concept of a git pre-commit hook and the tool pre-commit are two different things.

And/or: it would be nice to have instructions for how to run the required tools to make the CI checks happy that doesn't involve tools calling tools calling tools.

If you want an easy workflow, everything you need should be described at https://data-apis.org/array-api-extra/contributing.html#development-workflow. It is effectively:

1. install Pixi
2. pixi run lint, or the other commands documented there.

It can't get much simpler than that (this is a huge improvement on many projects I have contributed to). I am keen to make this workflow the best it can be, so let me know if you have any ideas for improvements.

If you do not want to install Pixi, or otherwise want to use a different workflow, that is fine, and I am keen to accept contributions which accommodate for alternative workflows, while not being a detriment to other workflows.

I will not be volunteering my time to work on speculative accommodations for alternative workflows, though, as I strongly believe that the Pixi workflow is both essential to and sufficient for the project.

With the changes to the liniting/pre-commit setup: is evilmartians/lefthook the lefthook you are now using?

Yes

This isn't specific to pixi, but more because five different repos use six different of these tools :-/

I feel your pain, which is why I am so excited about Pixi, as a potential replacement for all of them.

I added a delegation to dask to solve some of the errors related to take_along_axis, and because it is what we want because it implements quantile (more delegation to torch, cupy, etc to follow). However now a basic test fails. We can reproduce this with just basic dask:

import array_api_compat.dask.array as xp

x = xp.asarray([1, 2, 3, 4, 5])
xp.quantile(x, 0.5)

this leads to the following error:

File ~/miniconda/envs/array-api-extra/lib/python3.13/site-packages/dask/array/reductions.py:1567, in quantile(a, q, axis, out, overwrite_input, method, keepdims, weights, interpolation)
   1562 """
   1563 This works by automatically chunking the reduced axes to a single chunk if necessary
   1564 and then calling ``numpy.quantile`` function across the remaining dimensions
   1565 """
   1566 if axis is None:
-> 1567     if any(n_blocks > 1 for n_blocks in a.numblocks):
   1568         raise NotImplementedError(
   1569             "The da.quantile function only works along an axis.  "
   1570             "The full algorithm is difficult to do in parallel"
   1571         )
   1572     else:

File ~/miniconda/envs/array-api-extra/lib/python3.13/site-packages/dask/array/reductions.py:151, in any(a, axis, keepdims, split_every, out)
    149 @derived_from(np)
    150 def any(a, axis=None, keepdims=False, split_every=None, out=None):
--> 151     return reduction(
    152         a,
    153         chunk.any,
    154         chunk.any,
    155         axis=axis,
    156         keepdims=keepdims,
    157         dtype="bool",
    158         split_every=split_every,
    159         out=out,
    160     )

File ~/miniconda/envs/array-api-extra/lib/python3.13/site-packages/dask/array/_reductions_generic.py:141, in reduction(x, chunk, aggregate, axis, keepdims, dtype, split_every, combine, name, out, concatenate, output_size, meta, weights)
     41 """General version of reductions
     42
     43 Parameters
   (...)    138
    139 """
    140 if axis is None:
--> 141     axis = tuple(range(x.ndim))
    142 if isinstance(axis, Integral):
    143     axis = (axis,)

AttributeError: 'generator' object has no attribute 'ndim'

What is the normal procedure for dealing with something like this? Can we skip the test for dask?

What is the normal procedure for dealing with something like this? Can we skip the test for dask?

Yep, the normal procedure would be:

• open an upstream issue
• add a skip or xfail linking to the upstream issue, like

  array-api-extra/tests/test_funcs.py

  Lines 1047 to 1057 in 83897e5

  	@pytest.mark.xfail_xp_backend(Backend.SPARSE, reason="no argsort")
  	@pytest.mark.skip_xp_backend(Backend.ARRAY_API_STRICTEST, reason="no unique_values")
  	class TestSetDiff1D:
  	@pytest.mark.xfail_xp_backend(Backend.DASK, reason="NaN-shaped arrays")
  	@pytest.mark.xfail_xp_backend(
  	Backend.TORCH, reason="index_select not implemented for uint32"
  	)
  	@pytest.mark.xfail_xp_backend(
  	Backend.TORCH_GPU, reason="index_select not implemented for uint32"
  	)
  	def test_setdiff1d(self, xp: ModuleType):

This isn't specific to pixi, but more because five different repos use six different of these tools :-/

I feel your pain, which is why I am so excited about Pixi, as a potential replacement for all of them.

Can I install pixi in an existing conda env and it won't touch or modify things outside of that conda env? I don't want to blindly install something that (potentially) messes with my shell or might end up breaking things for other projects. And figuring out the answer to that question isn't easy, I think. Especially because most of the instructions are "just install pixi" - which if I was starting from scratch would be fine, but I want the "install pixi the hard way" instructions so I know what it might (or might not) break

Can I install pixi in an existing conda env and it won't touch or modify things outside of that conda env?

I don't think that is possible, but I will ask and let you know if it is.

I don't want to blindly install something that (potentially) messes with my shell or might end up breaking things for other projects. And figuring out the answer to that question isn't easy, I think.

The best I can do is say that I have had no such problems 🤷‍♂️