Unable to use an internal RCCL build

[gsitaram]

Setting the following. env. variables does not result in the provided RCCL lib and include file to be used in the COSMA build:

        export RCCL_ROOT_DIR="/path/to/install/rccl"
        export RCCL_INCLUDE_DIR="/path/to/install/rccl/include"
        export RCCL_LIB_DIR="/path/to/install/rccl/lib"

where the install directory has the following:

$ ls -R install/rccl
install/rccl:
include  lib

install/rccl/include:
rccl.h

install/rccl/lib:
librccl.so  librccl.so.1  librccl.so.1.0

[kabicm]

Thanks Gina!

@teonnik since you are a cmake expert, do you know how this can be fixed? I tried, but am certainly not a cmake ninja :D

[kabicm]

I now fixed the environment variables properly being passed to cmake in the latest commit in branch v2.6.0, with:

COSMA/cmake/FindRCCL.cmake

Line 18 in d6ca0ad

set(RCCL_ROOT_DIR $ENV{RCCL_ROOT_DIR})

and similar.

As @gsitaram said, it seems now it finds the include header files, but doesn't find the library. E.g. it finds the default librccl.so, but not the one that is specified by RCCL_LIB_DIR.

[kabicm]

@gsitaram Hey Gina, is this resolved?

[gsitaram]

Hi @kabicm , unfortunately, not yet. Would be great if we could get some help from some CMake experts. My attempts have been futile.