Prebuilt PyTorch Packages for IBM Z and LinuxONE

This is IBM CODAIT’s PyTorch prebuilt packages for IBM Z and LinuxONE. Our build is currently minimal and supports CPU only. This project is still in its beta phase, and any feedback is welcome.

Prerequisites

A GNU/Linux distribution with glibc >= 2.28. This includes RHEL 8 and Ubuntu 20.04.

You also need BLAS and LAPACK. On RHEL, install them with

$ sudo dnf install blas lapack

On Ubuntu, install them with

$ sudo apt install libblas3 liblapack3

Additionally, building the dependency NumPy (which will be automatically executed when installing PyTorch) requires the Python development files, a C compiler, and a Fortran compiler.

Installation

To install the latest package, for Python 3.6, run:

$ pip install https://github.com/CODAIT/pytorch-on-z/raw/packages/torch-1.8.1-cp36-cp36m-linux_s390x.whl

For Python 3.8, run:

$ pip install https://github.com/CODAIT/pytorch-on-z/raw/packages/torch-1.8.1-cp38-cp38-linux_s390x.whl

Known Issues

Not all tests for the C++ library can pass.

How We Built These Packages

We built the packages as follows on RHEL 8 on LinuxONE.

  1. Create an empty virtual environment and activate it:

     $ python3.8 -m venv ~/.pytorch-build  # Replace python3.8 with python3.6 if building for Python 3.6
     $ . ~/.pytorch-build/bin/activate
    
  2. Clone PyTorch source code and check out the PyTorch version to build:

     $ git clone --recursive https://github.com/pytorch/pytorch.git
     $ cd pytorch
     $ git checkout v1.8.1
     $ git submodule update --recursive
    
  3. Install dependencies:

     $ pip install -r requirements.txt
     $ pip install wheel
    
  4. Run the build commands:

     $ DEBUG=0 CMAKE_EXPORT_COMPILE_COMMANDS=YES USE_OPENMP=0 USE_DISTRIBUTED=0 USE_MKLDNN=0 USE_NCCL=0 USE_CUDA=0 USE_PYTORCH_QNNPACK=0 USE_XNNPACK=0 USE_QNNPACK=0 USE_NNPACK=0 USE_QNNPACK=0 USE_FBGEMM=0 BUILD_TEST=1 USE_NUMA=OFF python3 setup.py develop --cmake-only
     $ MAX_JOBS=1 python setup.py develop
    
  5. Run the tests:

     $ python test/run_test.py
    
  6. Create the package:

     $ python setup.py bdist_wheel
    

Now the package should be available in the directory dist/.