[Build] Make nccl optional (#3056)

* fix

* remove nvidiasmi

* fix

* fix docs

* fix

* fix

* 1

* fix

* remove

* skip deprecated kernel

* fix

* Revert "skip deprecated kernel"

This reverts commit c5ceb7f60dbbaf065b81cc3680757fd611d90ad3.

* fix

CMakeLists.txt

@@ -23,6 +23,7 @@ endif()
 # and add set(OPTION VALUE) to override these build options.
 # Alernatively, use cmake -DOPTION=VALUE through command-line.
 dgl_option(USE_CUDA "Build with CUDA" OFF)
+dgl_option(USE_NCCL "Build with NCCL support" OFF)
 dgl_option(USE_SYSTEM_NCCL "Build using system's NCCL library" OFF)
 dgl_option(USE_OPENMP "Build with OpenMP" ON)
 dgl_option(USE_AVX "Build with AVX optimization" ON)
@@ -159,21 +160,23 @@ list(APPEND DGL_SRC ${DGL_SRC_1})
 if(USE_CUDA)
   dgl_config_cuda(DGL_CUDA_SRC)
   list(APPEND DGL_SRC ${DGL_CUDA_SRC})
-
-  if (USE_SYSTEM_NCCL)
-    include(cmake/util/FindNccl.cmake)
-    include_directories(${NCCL_INCLUDE_DIR})
-  else()
-    include(cmake/modules/NCCL.cmake)
-    cuda_include_directories(BEFORE ${NCCL_INCLUDE_DIR})
-  endif()
+  if(USE_NCCL)
+    add_definitions(-DDGL_USE_NCCL)
+    if (USE_SYSTEM_NCCL)
+      include(cmake/util/FindNccl.cmake)
+      include_directories(${NCCL_INCLUDE_DIR})
+    else()
+      include(cmake/modules/NCCL.cmake)
+      cuda_include_directories(BEFORE ${NCCL_INCLUDE_DIR})
+    endif()
+  endif(USE_NCCL)
 
   list(APPEND DGL_LINKER_LIBS ${NCCL_LIBRARY})
 endif(USE_CUDA)
 
 if(USE_CUDA)
   cuda_add_library(dgl SHARED ${DGL_SRC})
-  if (NOT USE_SYSTEM_NCCL)
+  if (USE_NCCL AND NOT USE_SYSTEM_NCCL)
     add_dependencies(dgl nccl_external)
   endif()
 else(USE_CUDA)

src/runtime/cuda/nccl_api.cu

@@ -4,6 +4,7 @@
  * \brief Implementation of wrapper around NCCL routines.
  */
 
+#ifdef DGL_USE_NCCL
 
 #include "nccl_api.h"
 
@@ -782,5 +783,5 @@ DGL_REGISTER_GLOBAL("cuda.nccl._CAPI_DGLNCCLSparseAllToAllPull")
 }  // namespace runtime
 }  // namespace dgl
 
-
+#endif
 

tests/scripts/build_dgl.sh

@@ -14,7 +14,7 @@ CMAKE_VARS="-DBUILD_CPP_TEST=ON -DUSE_OPENMP=ON -DBUILD_TORCH=ON"
 CMAKE_VARS="$CMAKE_VARS -DTORCH_PYTHON_INTERPS=/opt/conda/envs/pytorch-ci/bin/python"
 
 if [ "$1" == "gpu" ]; then
-    CMAKE_VARS="-DUSE_CUDA=ON $CMAKE_VARS"
+    CMAKE_VARS="-DUSE_CUDA=ON -DUSE_NCCL=ON $CMAKE_VARS"
 fi
 
 if [ -d build ]; then