Consider packaging and distribution of DESC on conda-forge

[matthewfeickert]

In addition to distributing desc-opt on PyPI, it would be useful to have multi-platform built distributions distributed on conda-forge. This would allow for people to install in a self contained and isolated environment in user space through Pixi (e.g. pixi add desc).

This would mostly just require making a PR to https://github.com/conda-forge/staged-recipes and then following a recipe for an existing Python-with-compiled-extensions project like https://github.com/conda-forge/boost-histogram-feedstock.

I would be happy to help with the PR as well as answering any conda-forge related questions people might have.

[YigitElma]

Hi @matthewfeickert! Thanks for the suggestion. I am not very familiar with conda-forge, as far as I know it makes installation of compiled codes easier but DESC is pure Python, so users don't need to compile anything. Our biggest dependency jax is on conda-forge I think, but I am not sure how up to date it is.

Can you give a bit more information on what is required? What will be the maintanence cost? How will the GPU support work?

Thanks!

[matthewfeickert]

but DESC is pure Python, so users don't need to compile anything.

@YigitElma That's great (I was opening a few of these issues yesterday between meetings and didn't do a great job of checking the state of the project). For pure-Python projects these can be readily added as noarch:python projects as long as all of their dependencies exist on conda-forge already. You can do this by the following:

• Forking https://github.com/conda-forge/staged-recipes
• Making a new branch
• Installing Pixi
• Running (at the top level of the staged-recipes repo)

pixi run pypi desc-opt

• Submitting the results in a PR

Our biggest dependency jax is on conda-forge I think, but I am not sure how up to date it is.

As of recently (there were some build issues with jaxlib) jax on conda-forge is up to date with PyPI (okay, well technically one patch version behind but that's in progress):

$ pixi search jax
Using channels: conda-forge
jax-0.9.1-pyhd8ed1ab_0
----------------------

Name                jax
Version             0.9.1
Build               pyhd8ed1ab_0
Size                1.92 MiB
License             Apache-2.0
Timestamp           2026-03-25 16:38:24 UTC
Subdir              noarch
NoArch              python
File Name           jax-0.9.1-pyhd8ed1ab_0.conda
URL                 https://conda.anaconda.org/conda-forge/noarch/jax-0.9.1-pyhd8ed1ab_0.conda
MD5                 544954350739e4ef0c13999583f45ebc
SHA256              27b97dd0c20a9596e56d556bd3a28395c7195662774767f3c285bf248bb240e0

Dependencies:
 - importlib-metadata >=4.6
 - jaxlib >=0.9.1,<=0.9.1
 - ml_dtypes >=0.5.0
 - numpy >=2.0
 - opt_einsum
 - python >=3.11
 - scipy >=1.13
...

Can you give a bit more information on what is required? What will be the maintenance cost?

For most situations with pure-Python conda-forge packages ("noarch" in the conda world parlance) there is very little work. Conda-forge can use the sdist on PyPI as the source for the build and has https://github.com/regro-cf-autotick-bot which watches all "feedstock" repositories and continually checks their upstream sources for new releases. When one is found, regro-cf-autotick-bot automatically creates a PR to the feedstock that updates the version number and digest and attempts to build the new release of the package. For a pure-Python package the only reason this would fail would be if there is a dependency change in the release that needs to be reflected in the dependency manifest in the "recipe" of if a lower bound Python version has been set and the new release requires this to be bumped up. In both situations the error messages are transparent and this can usually be fixed through the GitHub web UI. While I do not recommend this, as I want to double check all releases, you have the ability to also have feedstocks have regro-cf-autotick-bot PRs be automatically merged without feedstock maintainer review if they pass.

How will the GPU support work?

jaxlib on conda-forge has build variants for CPU and GPU that can be either explicitly requested or picked up by configuration options. For example, if I tell Pixi that I want support for systems with CUDA v12.x on them, the CUDA-enabled builds of jaxlib will be resolved

$ pixi init /tmp/example && cd /tmp/example
✔ Created /tmp/example/pixi.toml
$ pixi workspace system-requirements add cuda 12
$ pixi add --no-install jax
✔ Added jax >=0.9.1,<0.10
$ pixi list jax
Name    Version  Build                            Size  Kind   Source
jax     0.9.1    pyhd8ed1ab_0                 1.92 MiB  conda  https://conda.anaconda.org/conda-forge
jaxlib  0.9.1    cuda129_py314hfcd48a0_200  175.40 MiB  conda  https://conda.anaconda.org/conda-forge

(here I'm on a machine with NVIDIA drivers, but you can spoof this with CONDA_OVERRIDE_CUDA).

In your situation,

DESC/requirements.txt

Line 1 in f8c0ec0

jax >= 0.5.0, != 0.5.1, != 0.5.2, != 0.6.0, != 0.7.1, < 0.10.0

the same thing would happen, so you basically get excellent CUDA support for free.

[YigitElma]

Thanks for the detailed explanation @matthewfeickert.

For a pure-Python package the only reason this would fail would be if there is a dependency change in the release that needs to be reflected in the dependency manifest in the "recipe"

Unfortunately, due to many breaking changes in jax, we constantly update our requirements file. So, this might be more work and testing than we can spend time on right now.

Currently, we don't plan to work on this, given the busy schedule of other projects. However, if you are interested, you are welcome to create a conda-forge distribution as a community-supported version of DESC.

[matthewfeickert]

For a pure-Python package the only reason this would fail would be if there is a dependency change in the release that needs to be reflected in the dependency manifest in the "recipe"

Unfortunately, due to many breaking changes in jax, we constantly update our requirements file. So, this might be more work and testing than we can spend time on right now.

A conda package is not an independent system from the underlying library — it is a distribution of the library — so there really isn't independent "testing". In the situation in which you've changed the libraries dependencies in a release then you just update those same dependencies in the automatically opened PR from https://github.com/regro-cf-autotick-bot.

you are welcome to create a conda-forge distribution as a community-supported version of DESC.

👍 @YigitElma do you or any other DESC maintainers want to be listed as maintainers, with the understanding that you aren't going to do help out, so that you retain ownership control over the distribution of your software?

[YigitElma]

so there really isn't independent "testing"

I meant testing of the installation on different platforms.

do you or any other DESC maintainers want to be listed as maintainers

Sure, you can add me as a maintainer. @f0uriest @ddudt @dpanici would you like to be added as maintainers too?

[matthewfeickert]

I meant testing of the installation on different platforms.

Ah, conda-forge handles this all automatically. It builds the package and then the test sections of the recipe will install the built package as a user would in a new environment and then execute all the tests you define before a build can be published. Though, in fairness, what I described covers everything on platform-specific builds. desc-opt would be a "noarch" project and as those are supposed to be pure-Python and so agnostic to the platform, and so only built and tested on Linux (though all of the platform-specific dependency builds do get tested on their own feedstocks).

Sure, you can add me as a maintainer.

Awesome and thanks. 👍 I hope to get this up soon (I was wrong though and there are two dependencies not yet on conda-forge that I need to get up first).

[matthewfeickert]

DESC has been added to conda-forge as desc-opt in conda-forge/staged-recipes#32937 (@f0uriest orthax was as well as it is a required dependency, c.f. f0uriest/orthax#91) and exists as a feedstock at https://github.com/conda-forge/desc-opt-feedstock.

Name	Downloads	Version	Platforms

This presumed that PRs #2163 and #2160 would be accepted and patches them in. If PR #2160 does require additional patches following a second maintainer review those can be patched into the conda-forge feedstock and rebuilt.

If the maintainer team is interested in validating behavior or trying the distribution out on a Linux machine I'd suggest:

• Installing Pixi
• Creating a Pixi workspace

pixi init example && cd example

• Installing desc-opt

pixi add desc-opt

• Running commands in the environment with Pixi

pixi run <commands>

or creating an interactive subshell with the environment activated

pixi shell

If it is of interest, I can PR additions of installation instructions from conda-forge.

---

Example run on my laptop:

$ pixi init example && cd example
$ pixi workspace system-requirements add cuda 12
$ pixi add desc-opt
$ pixi run python -c 'import desc; from desc.backend import print_backend_info; print_backend_info()'
DESC version=0.17.1.
Using JAX backend: jax version=0.9.2, jaxlib version=0.9.2, dtype=float64.
Using device: CPU, with 20.01 GB available memory.
$ pixi run python -c 'import desc; from desc import set_device; set_device("gpu"); from desc.backend import print_backend_info; print_backend_info()'
DESC version=0.17.1.
Using JAX backend: jax version=0.9.2, jaxlib version=0.9.2, dtype=float64.
Using device: NVIDIA GeForce RTX 4060 Laptop GPU (id=0), with 7.99 GB available memory.

[YigitElma]

@matthewfeickert, since #2163 and #2160 are merged, I assume this is resolved?

[matthewfeickert]

I assume this is resolved?

@YigitElma From a backporting patches perspective yes, but do you want documentation noting its existence?

If it is of interest, I can PR additions of installation instructions from conda-forge.

Also ping RE:

@f0uriest @ddudt @dpanici would you like to be added as maintainers too?

[YigitElma]

@matthewfeickert If you would like, you can create a PR to update the installation instructions. I would suggest having a separate section before this line, it can be community-supported conda-forge or something on that front. As I noted before, we currently don't have enough bandwidth to support the conda-forge distribution, so it would be nice if you could include your contact details (otherwise, I don't think other developers want to advertise the option on our official docs).

Also ping RE:

I talked to them in person, and they are not interested in being added as maintainers right now.