RDMA over Thunderbolt

[okuvshynov]

Hi folks,

According to

https://developer.apple.com/documentation/macos-release-notes/macos-26_2-release-notes#RDMA-over-Thunderbolt

there seem to be some RDMA implementation which I suppose doesn't involve network stack and allows to get low latency communication (potentially very useful for expert parallelism/tensor parallelism for llm inference).

Are there any more details on that? Maybe some estimates/benchmark results for roundtrip for several KB of data (~activation size)?

Any API I could experiment on?

Thank you!

[awni]

This PR has some benchmarks, examples of how to use the new API etc: #2808

[okuvshynov]

Thank you!