Allowing SPM to compile on Linux with CUDA

[gmondada]

Proposed changes

This PR introduces an SPM plug-in allowing CUDA to be compiled by SPM on Linux.

Example1 can be compiled on Linux with command:

swift build --product Example1

SPM invokes nvcc, therefore you need the CUDA toolkit to be installed.

A few notes:

• You need at least CUDA 13.2. Version 12 has a very limited support for clang.
• This solution rely on “CGen” SPM experimental feature, which is only available in Swift 6.3.
• Code generated at compile time is managed by tools/update-mlx.sh.
• On my computer I need to compile with --jobs=2 for limiting the memory required during compilation. Solutions to mitigate this will be proposed in further PRs.

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[gmondada]

@davidkoski @Joannis 👆

[gmondada]

To compile with CUDA on Mac, just create a docker with https://github.com/gmondada/cuda-toolkit-docker/blob/main/Dockerfile.ubuntu24

[davidkoski]

Xcode 26.3
...
package 'mlx-swift' is using Swift tools version 6.3.0 but the installed version is 6.2.4

Let me see if we have Xcode 26.4 available in CI.

[davidkoski]

See #414 -- it looks like we can use the latest or a particular version of Xcode:

        env:
          DEVELOPER_DIR: /Applications/Xcode-latest.app

this should be done for all the steps that do xcodebuild or xcrun