Skip to content

INSTALL.md

Latest commit

 

History

History
121 lines (96 loc) · 3.52 KB

INSTALL.md

File metadata and controls

121 lines (96 loc) · 3.52 KB

INSTALLATION

For now, this version has only been tested on Linux with NVIDIA GPUs that support CUDA 12.x.

PREREQUISITES

  • Run git clone --recursive https://github.com/apache/incubator-mxnet mxnet
  • Install pre-requisites:
$ sudo apt-get update
$ sudo apt-get install -y build-essential git ninja-build ccache \
    libopenblas-dev libopencv-dev cmake gfortran htop
$ sudo apt-get -y install gcc-12 g++-12 gfortran-12
  • If you have NVIDIA GPUs installed on the Linux system, install CUDA 12.6 SDK
  • The older CUDA 12.2 SDK is not supported at the moment due to header file issues.
$ which nvcc
/usr/local/cuda/bin/nvcc

PATCH HEADER FILES IN CUDA

Patch CUDA header file in CUDA 12.6 SDK /usr/local/cuda/include/cub//device/dispatch/dispatch_histogram.cuh and change all invocations of cuda:: to look like ::cuda::, to avoid the compiler producing ambiguous errors. In CUDA 12.6 SDK this will be in line 422.

In CUDA 12.2 SDK this will be in several lines.

Reported issue: NVIDIA/cccl#2939

The reason this is done is that c++11 compilers like gcc-11 or gcc-12 seems to confuse itself for the ::cuda namespace and the various sub-namespaces also named cuda in the MXNet repository. Fixing this one header file solves the issues with compiling. You only need to do this if your CUDA SDK version is between 12.0 and 12.6. We believe that based on the above reported issue and the fix, it will not be necessary in CUDA SDK 12.7 or greater.

CPU ONLY BUILD

  • Run cmake with accceptable defaults. For pure CPU-only computations, you have to disable CUDA support using the USE_CUDA option.
$ mkdir -p _build_cpu
## or use your own custom install directory like /usr/local or $HOME/local
$ export CMAKE_INSTALL_PREFIX=${HOME}/local/cpu
$ mkdir -p $CMAKE_INSTALL_PREFIX
$ cd _build_cpu
$ cmake -DUSE_CPP_PACKAGE=1 -DUSE_CUDA=OFF -DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX} ..
$ make -j4
$ make install
$ cd ..

CUDA BUILD

  • Run cmake with accceptable defaults, and set USE_CUDA to ON.
$ mkdir -p _build_cuda
## or use your own custom install directory like /usr/local or $HOME/local
$ export CMAKE_INSTALL_PREFIX=${HOME}/local/mxnet/cuda
$ mkdir -p $CMAKE_INSTALL_PREFIX
$ cd _build_cuda
$ CC=gcc-12 CXX=g++-12 FC=gfortran-12 cmake \
    -DUSE_CPP_PACKAGE=1 -DUSE_CUDA=ON \
    -DMXNET_CUDA_ARCH=7.5 \
    -DCMAKE_CUDA_ARCHITECTURES=sm_75 \
    -DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX} ..
$ make -j4
$ make install
$ cd ..

RECOMMENDED TOOLS

To check if the CUDA-enabled GPU is being used, we recommend you install nvtop.

$ sudo apt -y install libdrm-dev libsystemd-dev libudev-dev libncurses-dev \
    libncursesw5-dev git cmake build-essential
$ git clone https://github.com/Syllo/nvtop
$ cd nvtop
$ mkdir -p Release
$ cd Release
###  you can set this to /usr/local or $HOME/local or something you like
$ export CMAKE_INSTALL_PREFIX=$HOME/local/
$ cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}..
$ make -j2
## if you're using $HOME/local
$ make install
## else if you're using /usr/local
$ sudo make install
### end if
$ export PATH=$CMAKE_INSTALL_PREFIX/bin:$PATH
$ which nvtop
/home/<username>/local/bin/nvtop

We recommend building nvtop from the source since you get the latest updates.

DATA DOWNLOADS

  • Download sample training data into a folder named data
$ ./cpp-package/example/get_data.sh
  • Using MXNET in C++. Set the header file and the LD_LIBRARY_PATH correctly
#include <mxnet-cpp/MxNetCpp.h>