It is actually simple to get Blender + Cycles + Nvidia/CUDA working on Mac OS X 10.13.6 (High Sierra). It takes 5 steps assuming Xcode is already installed:

• Install Nvidia CUDA toolkit 10.2 as per Bruno Wego's method.

  
    wget 'https://developer.download.nvidia.com/compute/cuda/10.2/Prod/local_installers/cuda_10.2.89_mac.dmg' && \
    hdiutil attach cuda_10.2.89_mac.dmg \
        -nobrowse \
        -mountpoint \
        /Volumes/CUDAMacOSXInstaller
  
    open /Volumes/CUDAMacOSXInstaller/CUDAMacOSXInstaller.app
  
    hdiutil detach /Volumes/CUDAMacOSXInstaller && rm ./cuda_10.2.89_mac.dmg
    

• Prepare for a build on mac os x as per Blender instructions, quick setup. Do the

  make update
  make

  to test the ability to build the CPU version first (and fix whatever is missing until the build works).
• Fix file blender/intern/cycles/device/cuda/device_cuda_impl.cpp, in function CUDADevice::compile_kernel_get_common_cflags. Add the string "-std=c++14 " as first argument of the string_printf (don't omit the trailing space!). It should look like this:

  
       string cflags = string_printf(
              "-std=c++14 "
              "-m%d "
              "--ptxas-options=\"-v\" "
              "--use_fast_math "
              "-DNVCC "
              "-I\"%s\"",
              machine,
              include_path.c_str());
    

• Modify the file lib/darwin/nanovdb/include/nanovdb/NanoVDB.h (that's obtained during the make update), fixing all the static_assert calls for which the 1st param is a C++ templatised value that have an extra comma by simply putting the expression in parenthesis, so the non-C++14 compiler don't get confused. This is the diff I got after the fix:

  
        887c887
        <     static_assert(is_same<int, typename CoordT::ValueType>::value, "Expected \"int\" coordinate type");
        ---
        >     static_assert((is_same<int, typename CoordT::ValueType>::value), "Expected \"int\" coordinate type");
        1515,1516c1515,1516
        <     static_assert(is_same<NodeType<0>, LeafNodeType>::value, "NodeType<0> error");
        <     static_assert(is_same<NodeType<3>, RootType>::value, "NodeType<3> error");
        ---
        >     static_assert((is_same<NodeType<0>, LeafNodeType>::value), "NodeType<0> error");
        >     static_assert((is_same<NodeType<3>, RootType>::value), "NodeType<3> error");
        1594c1594
        <     static_assert(is_same<NodeType<NodeT::LEVEL>, NodeT>::value, "Tree::getNode: unvalid node type");
        ---
        >     static_assert((is_same<NodeType<NodeT::LEVEL>, NodeT>::value), "Tree::getNode: unvalid node type");
        1611c1611
        <     static_assert(is_same<NodeType<NodeT::LEVEL>, NodeT>::value, "Tree::getNode: unvalid node type");
        ---
        >     static_assert((is_same<NodeType<NodeT::LEVEL>, NodeT>::value), "Tree::getNode: unvalid node type");
        2425c2425
        <         static_assert(is_same<T, NodeT>::value, "ReadAccessor::getNode: Invalid node type");
        ---
        >         static_assert((is_same<T, NodeT>::value), "ReadAccessor::getNode: Invalid node type");
    

• Build blender with GPU support for cycles (replace sm_61 by the CUDA Compute level your card supports):

  
    make WITH_CYCLES_CUDA_BINARIES=ON DCYCLES_CUDA_BINARIES_ARCH=sm_61
    

I used the latest git version (commit fa82a15676eb37a7e73d1e3a0e8095684842376d, version 2.92) to build, my system is an aging hackintosh running Xcode 10.1 and WebDrives 387.10.10.10.40.132 (ie, about a year old) with a GTX 1050 Ti and yet the whole thing work just fine.
It's great to be back!