🚀 The feature, motivation and pitch

I want to dynamically set the number of SMs in torch.distributed.all_reduce. NCCL supports using the nccl_max_nchannels environment variable setting.but cant dynamically set in the program. It is mentioned here that ncclCommInitRankConfig can be used in the program (link),, but the corresponding setting is not found in torch. Can this capability be supported? This is useful in inference optimization scenarios

Alternatives

No response

Additional context

No response

cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o