HPC IDUN Cluster

How to build Pytorch 1.3.0 for the IDUN cluster

This blog post noticed remarkable performance gains from building pytorch from source.

To build it from source on the IDUN cluster was kind of a hassle, but this script should do it:

module load GCC/8.2.0-2.31.1
module load CUDA/10.1.105
module load Anaconda3/2018.12
module load cuDNN/7.4.2.24
module load NCCL/2.4.2

conda create --name py37 python=3.7
# Got error message when running activate for the first time.
# Had to do what it said, and log onto the node again
conda activate py37

# pytorch dependencies
conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing --yes
conda install -c pytorch magma-cuda90 --yes
conda install opencv --yes


export CUDNN_INCLUDE_DIR="/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/cuDNN/7.4.2.24/include/"
export CUDNN_LIB_DIR="/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/cuDNN/7.4.2.24/lib64/"
export NCCL_INCLUDE_DIR="/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/NCCL/2.4.2/include/"
export NCCL_LIBRARY="/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/NCCL/2.4.2/lib/"
export NCCL_LIB_DIR="/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/NCCL/2.4.2/lib/"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/share/apps/software/Compiler/GCC-CUDA/8.2.0-2.31.1-10.1.105/NCCL/2.4.2/lib/"
export USE_CUDA=1 USE_CUDNN=1 USE_MKLDNN=1 USE_MKLDNN_CBLAS=1 USE_OPENCV=1

git clone --recursive https://github.com/pytorch/pytorch pytorch_source
cd pytorch_source
# Checkout to version 1.3.0
git checkout tags/v1.3.1
git submodule sync
git submodule update --init --recursive


# install pytorch
python setup.py install

cd ..
git clone https://github.com/pytorch/vision
cd vision
git checkout tags/v0.4.2

python setup.py install