In this file - https://github.com/pytorch/pytorch/blob/master/cmake/Modules/FindNCCL.cmake
Why aren’t we checking if nccl.h exists in the default location for unix - /usr/local/cuda
Or is it being done implicitly somewhere?
I think it makes sense to set NCCL_ROOT to /usr/loca/cuda if it is not already set since Nvidia recommends install nccl in that path.