strong compatibility to different python and pytorch versions

cuda/moe.cpp

@@ -116,7 +116,11 @@ class HackNCCLGroup: public c10d::ProcessGroupNCCL {
 public:
 	ncclComm_t getcomm(at::Device dev) {
 		auto key = std::to_string(dev.index());
+#ifdef ENABLE_NCCL_P2P_SUPPORT
 		auto v = getNCCLComm(key, {dev}, c10d::OpType::ALLTOALL);
+#else
+		auto v = getNCCLComm(key, {dev});
+#endif
 		if (v.size() == 0) {
 			std::cerr << "PyTorch has nothing\n";
 			return 0;

tests/test.sh

@@ -20,8 +20,12 @@ fi
 
 mkdir -p logs
 
+PYTHON_EXEC=python3
+PYTHON_VERSION=$($PYTHON_EXEC --version)
+PYTHON_REVISION=${PYTHON_VERSION:7:3}
+
 SCRIPT_PATH=$(dirname $(dirname $(realpath $0)))
-export PYTHONPATH=$SCRIPT_PATH:$SCRIPT_PATH/build/lib.linux-x86_64-3.7:$PYTHONPATH
-export LD_LIBRARY_PATH=/home/laekov/.local/lib/python3.7/site-packages/torch/lib:$LD_LIBRARY_PATH
+export PYTHONPATH=$SCRIPT_PATH:$SCRIPT_PATH/build/lib.linux-x86_64-$PYTHON_REVISION:$PYTHONPATH
+export LD_LIBRARY_PATH=/home/laekov/.local/lib/python$PYTHON_REVISION/site-packages/torch/lib:$LD_LIBRARY_PATH
 
-exec python3 $@ 2>logs/$RANK.log
+exec $PYTHON_EXEC $@ 2>logs/$RANK.log