cmake : support build for iOS/tvOS

[jhen0409]

Fix #3106.

Update the CMakeLists.txt and add CI steps to ensure to continue to work.

[alonfaraj]

Great!
Maybe it worth to have a separate job for each build (iOS, tvOS) and not putting all together inside macOS-latest-cmake?
It would be easier to monitor if one of them fails and keep them independent.

[jhen0409]

Great! Maybe it worth to have a separate job for each build (iOS, tvOS) and not putting all together inside macOS-latest-cmake? It would be easier to monitor if one of them fails and keep them independent.

Yeah it might be better and same as other builds, will change it before merge.

[cyrilzakka]

Would it be possible to add VisionOS support as well?

[jhen0409]

Would it be possible to add VisionOS support as well?

I've tried an iOS build on visionOS simulator before and it works well. Not yet to try with a visionOS project, but I think it should be easy by add the OS target.

[cyrilzakka]

Thanks! In case you also have a sample Swift/VisionOS project to look at, I'd love to take a look at it too! I tried using llm_farm which is based off of this project but it crashes on model load.

[jhen0409]

Thanks! In case you also have a sample Swift/VisionOS project to look at, I'd love to take a look at it too! I tried using llm_farm which is based off of this project but it crashes on model load.

It was an internal iOS project with target visionOS + Designed for iPad. I can switch back to Xcode beta to give it try again in the next few days.

Also, if it's a Simulator, during to Metal limitations we can't use constant buffers more than 14, so remember to use ngl = 0.

[cyrilzakka]

@jhen0409 Sorry to have to ask, but where do I set the ngl 0 flag in Swift? I see it as an option in the CI but not anywhere else (e.g. ggml-metal). Thanks for your patience!

[jhen0409]

@jhen0409 Sorry to have to ask, but where do I set the ngl 0 flag in Swift? I see it as an option in the CI but not anywhere else (e.g. ggml-metal). Thanks for your patience!

To be clear, it's called n_gpu_layers in the parameters.