2021-12-16 Build onnxruntime on WSL (Windows Linux Subsystem)#

I tried to build onnxruntime-training for GPU on WSL (Windows Linux Subsystem). I took the distribution Ubuntu 20.04. Paths should be updated according to your installation.

some useful commands once installed

nvidia-smi
nsys

Let’s assume WSL is installed, otherwise, here are some useful commands.

# see all local distributions
wsl -s -l

# see available distributions online
wsl --list --online

# install one distribution
wsl --install -d Ubuntu-20.04

Installation of required packages.

sudo apt-get install cmake
sudo apt-get install zlib1g-dev
sudo apt-get install libssl-dev
sudo apt-get install python3-dev
sudo apt-get install libhwloc-dev
sudo apt-get install libevent-dev

Installation of cmake.

curl -OL https://github.com/Kitware/CMake/releases/download/v3.22.1/cmake-3.22.1.tar.gz
tar -zxvf cmake-3.22.1.tar.gz
cd cmake-3.22.1
./bootstrap
make
sudo make install
export PATH=~/install/cmake-3.22.1/bin/:$PATH

Installation of openmpi:

gunzip -c openmpi-4.1.2.tar.gz | tar xf -
cd openmpi-4.1.2
./configure --prefix=/usr/local --with-cuda
make all install

Installation of CUDA (choose a compatible version with pytorch, 11.3 for example).

See CUDA on WSL User Guide

export CUDA_VERSION=11.3
export CUDA_VERSION_=11-3
wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-wsl-ubuntu.pin
sudo mv cuda-wsl-ubuntu.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/${CUDA_VERSION}.0/local_installers/cuda-repo-wsl-ubuntu-${CUDA_VERSION_}-local_${CUDA_VERSION}.0-1_amd64.deb
sudo dpkg -i cuda-repo-wsl-ubuntu-${CUDA_VERSION_}-local_${CUDA_VERSION}.0-1_amd64.deb
sudo apt-key add /var/cuda-repo-wsl-ubuntu-${CUDA_VERSION_}-local/7fa2af80.pub
sudo apt-get update
sudo apt-get -y install cuda

Installation of cudnn (after it is downloaded):

sudo dpkg -i cudnn-local-repo-ubuntu2004-8.3.1.22_1.0-1_amd64.deb
sudo apt-key add /var/cudnn-local-repo-*/7fa2af80.pub
sudo apt-get update
sudo apt-get install libcudnn8
sudo apt-get install libcudnn8-dev

Installation of nccl

See Install NCCL.

sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/7fa2af80.pub
sudo add-apt-repository "deb https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/ /"
sudo apt update
sudo apt install libnccl2 libnccl-dev

Installation of pytorch:

python3 -m pip install torch==1.10.1+cu113 torchvision==0.11.2+cu113 torchaudio==0.10.1+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Then to check CUDA is available:

import torch
print(torch.cuda.is_available())

Build onnxruntime-training:

alias python=python3
export CUDA_VERSION=11.3
export CUDACXX=/usr/local/cuda-${CUDA_VERSION}/bin/nvcc
export MPI_HOME=~/install/openmpi-4.1.2
python3 ./tools/ci_build/build.py --skip_tests --build_dir ./build/linux_gpu --config Release --use_mpi false --enable_training --enable_training_torch_interop --use_cuda --cuda_version=${CUDA_VERSION} --cuda_home /usr/local/cuda-${CUDA_VERSION}/ --cudnn_home /usr/local/cuda-${CUDA_VERSION}/ --build_wheel --parallel

Option --parallel 1 can be used to fix the parallelism while building onnxruntime. Option –use_mpi false can be replaced by –mpi_home /usr/local/lib/openmpi.

Another option is to use a docker: Running Existing GPU Accelerated Containers on WSL 2.