Peter Andreas Entschev

https://github.com/cupy/cupy/pull/6908 is being planned for CuPy v13, that will make `import cupy` lazy-loading, which should prevent this memory spike.

> It might also be nice to update the High Performance Network section to talk about UCX, but I could only find NVIDIA/GPU specific examples. It would be good to...

> @pentschev there are definitely folks out there using IPoIB as many docs pages and examples show how to set your interface to the IB one. I just wondered what...

@vyasr could you comment here? I think you know how to fix it and have a proposal for that.

@yosefe @brminich could you comment whether this is an acceptable solution for you or if there are reasons to have things done differently?

> is #9696 it needed in previous versions? Thanks @tvegas1 ! Yes, could it be backported to 1.16? That would simplify a lot for us in the near-term.

Thanks @tvegas1 !

> @pentschev I recall you added the support for `CUPY_EXPERIMENTAL_SLICE_COPY` in #2079. Do you recall why we didn't also allow using ellipsis here? I think we never discussed using ellipsis...

Thanks @revans2 and @abellina for the comments. That was indeed one of our ideas, just use a bounce buffer for UCX, and the remaining memory for our regular pool. It...

I must say as much as I like the idea of packing things together, I'm a bit concerned with doubling the memory footprint. As we have already many memory pressing...