cuda-modules: fix and clean up multiplex builder package selection logic

[ConnorBaker]

This is a fix for a regression introduced in #368366, which manifests as cudnn (and likely other packages produced by the multiplex builder) selecting packages for the wrong architecture.

Things done

• Built on platform(s)
  • x86_64-linux
  • aarch64-linux
  • x86_64-darwin
  • aarch64-darwin
• For non-Linux: Is sandboxing enabled in nix.conf? (See Nix manual)
  • sandbox = relaxed
  • sandbox = true
• Tested, as applicable:
  • NixOS test(s) (look inside nixos/tests)
  • and/or package tests
  • or, for functions and "core" functionality, tests in lib/tests or pkgs/test
  • made sure NixOS tests are linked to the relevant packages
• Tested compilation of all packages that depend on this change using nix-shell -p nixpkgs-review --run "nixpkgs-review rev HEAD". Note: all changes have to be committed, also see nixpkgs-review usage
• Tested basic functionality of all binary files (usually in ./result/bin/)
• 25.05 Release Notes (or backporting 24.11 and 25.05 Release notes)
  • (Package updates) Added a release notes entry if the change is major or breaking
  • (Module updates) Added a release notes entry if the change is significant
  • (Module addition) Added a release notes entry if adding a new NixOS module
• Fits CONTRIBUTING.md.

---

Add a 👍 reaction to pull requests you find important.

[GaetanLepage]

Diff looks fine to me. Although I must admit that I am not familiar with this module, so consider it as an out-of-context sanity check.

[ConnorBaker]

Rebased on master to fix merge conflict since #405664 touched parts of the same file.

[ConnorBaker]

Merging since this allows building PyTorch.

[K900]

This broke eval of python312Packages.jax-cuda12-pjrt and possibly others on Darwin.

[WeetHet]

I don't see any filtering done in the previous version, so the issue is probably here

[SomeoneSerge]

Yes, iirc we switched from filtering the candidates to sorting them specifically as a dumb way to avoid missing attributes:

#282185
#276800

Been working offline for two days, need time to reload context, feel free to act ahead

[SomeoneSerge]

Could just special case the empty list branch

[SomeoneSerge]

This broke eval of python312Packages.jax-cuda12-pjrt and possibly others on Darwin.

Confirmed:

Details

❯ nix-instantiate --arg config '{ allowUnfree = true; }' -A python312Packages.jax-cuda12-pjrt --argstr system x86_64-darwin
       error: attribute 'cudnn' missing
       at /home/someone/Sources/nixpkgs/pkgs/development/python-modules/jax-cuda12-pjrt/default.nix:104:60:
          103|     # https://jax.readthedocs.io/en/latest/installation.html#pip-installation-nvidia-gpu-cuda-installed-locally-harder
          104|     broken = !(cudaAtLeast "12.1") || !(lib.versionAtLeast cudaPackages.cudnn.version "9.1");
             |                                                            ^
          105|   };

Odd that this passed the checks (CC @infinisil if you're still working on github workflows?)

[ConnorBaker]

Unless someone beats me to it I’ll make a PR to predicate broken on cudaSupport.

[SomeoneSerge]

predicate broken on cudaSupport.

Not sure I understand what's the idea?

[ConnorBaker]

Sorry that was unclear, I meant something like broken = config.cudaSupport && (<existing arguments OR’d together>).

Although, I’ll also see if there’s a simple way to have some sort of placeholder attribute in the CUDA package sets when none of the manifest entries correspond to the Nix system doing evaluation.

[SomeoneSerge]

something like broken = config.cudaSupport && (<existing arguments OR’d together>)

But where do you want to add that? EDIT: If you mean the python package, I think we should not do that. The bug is on cudaPackages side, and the error the python package should display is that it depends on a package unsupported by the platform IMO

have some sort of placeholder attribute in the CUDA package sets when none of the manifest entries correspond to the Nix system doing evaluation.

👍🏻 that's what we need to restore

[ConnorBaker]

See #406207 for a fix.