[FEATURE] Publish Vulkan & Metal binary wheels to PyPI

[iwr-redmond]

Feature Description

It would be helpful to consider publishing binary builds to PyPI for each version to supplement the existing source archive: Vulkan for Windows and Linux, and Metal for OSX.

This would greatly assist users who do not use the correct --index-url or --extra-index-url parameter when installing or upgrading the Nexa SDK package and experience various challenges as result.

The Windows and OSX builds already exist and it should be trivial to publish them to PyPI instead of just the source package. However, there is currently no Vulkan build for Linux.

Similar Features or References

• Vulkan performance statistics have recently been catalogued in llama.cpp#10879
• Nomic takes a similar approach to the one suggested here, building with Vulkan on Linux and Windows and offering Cuda as an extra feature.
• Both Upscayl and sd.cpp.gui.wx use Vulkan on all platforms, with the latter also offering a Cuda backend

Updates

02/11/2025: Correction regarding Vulkan, discovered while working on #386