Add recipe for mlx/metal

[hughjonesd]

I'm writing an R bridge to Apple's mlx library. Best practice seems to be to vendor mlx (it's MIT-licensed), and compile it when the package is installed. But doing so requires Apple's "metal" shader compiler, which is "only available in full Xcode not just Command Line Tools" (says my llm).

metal seems like something that R developers might be interested in, since it lets us use the GPU.

Can you advise whether either (a) you might produce a recipe for it or (b) you might install it, on CRAN macos servers or (c) you might install the mlx headers on CRAN (brew install mlx should suffice)?

[coatless]

@hughjonesd not sure CRAN will accept Rmlx as it would be limited to just Apple Silicone hardware and CRAN requires at least 2 major OSes being supported.

Package authors should make all reasonable efforts to provide cross-platform portable
code. Packages will not normally be accepted that do not run on at least two of the
major R platforms. Cases for Windows-only packages will be considered, but CRAN
may not be the most appropriate place to host them.

https://cloud.r-project.org/web/packages/policies.pdf#page=4

Like the initiative though and it's cool to see Claude doing this much heavy lifting on a GPU package. Any interest in joining for a Bandicoot wrapper instead?

[hughjonesd]

I think mlx can support a CUDA backend on linux, so that's my plan. That would give 2/3 platforms.

Haha yes, without them I'd be lost! codex is better but boasts less in commit messages.

[s-u]

I had a look at https://ml-explore.github.io/mlx/build/html/install.html and the requirements are way too high for CRAN at the moment (we require support for macOS 11 but it needs macOS 14+). We are considering macOS 14 as target for R 4.6.0 so it may become viable then, but for now you'll have to go with a source install.

[hughjonesd]

Ok. A source install will require Apple Metal, which I think is part of Xcode... might CRAN accept that, assuming it also builds on Linux (via either CPU/only or CUDA backends)?

[s-u]

I don't know - you'd have to ask. I have my doubts, because it's still now clear how it would look like - all the instructions are for Python modules. You can start looking into it and as I said, maybe ti will be mature enough when R 4.6.0 comes out so you can then put it on CRAN.