For PyTorch, I get this error
-- Could NOT find NCCL (missing: NCCL_INCLUDE_DIR NCCL_LIBRARY)
CMake Warning at third_party/gloo/cmake/Dependencies.cmake:96 (message):
Not compiling with NCCL support. Suppress this warning with
-DUSE_NCCL=OFF.
Call Stack (most recent call first):
third_party/gloo/CMakeLists.txt:56 (include)
Part of third_party/gloo/cmake/Dependencies.cmake
that contains the error is
if(USE_CUDA AND USE_NCCL)
# NCCL_EXTERNAL is set if using the Caffe2 bundled version of NCCL
if(NCCL_EXTERNAL)
include_directories(SYSTEM ${NCCL_INCLUDE_DIRS})
list(APPEND gloo_cuda_DEPENDENCY_LIBS ${NCCL_LIBRARIES} dl rt)
else()
find_package(nccl REQUIRED)
if(NCCL_FOUND)
include_directories(SYSTEM ${NCCL_INCLUDE_DIRS})
list(APPEND gloo_cuda_DEPENDENCY_LIBS ${NCCL_LIBRARIES} dl rt)
else()
message(WARNING "Not compiling with NCCL support. Suppress this warning with -DUSE_NCCL=OFF.")
set(USE_NCCL OFF)
endif()
endif()
endif()
Although, I have set env variables correctly, I would like to get more verbosity when cmake is reading that file. I mean, I want to know what is the value of NCCL_INCLUDE_DIRS when cmake is reading that file.
In the terminal, I see nothing wrong.
$ echo $NCCL_INCLUDE_DIRS
/opt/nccl-2.4.8/include
$ echo $NCCL_LIBRARIES
/opt/nccl-2.4.8/lib