Optional NCCL installation for PyPi wheel

[trivialfis]

Hi all, the XGBoost binary wheel size recently exceeded 200MB limit set by PyPi. I propose we remove the static linking of NCCL in our binary wheel to reduce the size. #7930 helped enable training distributed GPU without NCCL and as a result, XGBoost can work on MNMG without NCCL, albeit might suffer some performance loss. We can optionally emit a warning for users to install NCCL as a runtime dependency and dlopen the shared object when it's needed.

Alternative

• Compile our own nccl with a smaller set of supported archs. I removed the sm_35 and sm_50 from NCCL build, the striped static library is of size 190M build/lib/libnccl_static.a. I'm not sure if this can actually help in the long term.
• Reduce the number of GPU archs supported by XGBoost.
• Maybe submit a feature request for nccl to share a binary wheel on PyPi?

Related: cupy/cupy#4850

[trivialfis]

cc @RAMitchell @hcho3

[hcho3]

Why is this closed?

[trivialfis]

@RAMitchell worries that the performance hit might be too much and a simple warning is not sufficient for asking users to install nccl (which is indeed not an easy task for pip users).

[trivialfis]

We will have to find other options.