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/IBM/pytorch-on-z/raw/packages/torch-1.10.0-cp36-cp36m-linux_s390x.whl

For Python 3.8, run:

$ pip install https://github.com/IBM/pytorch-on-z/raw/packages/torch-1.10.0-cp38-cp38-linux_s390x.whl

For Python 3.9, run:

$ pip install https://github.com/IBM/pytorch-on-z/raw/packages/torch-1.10.0-cp39-cp39-linux_s390x.whl

Known Issues

Not all tests for the C++ library can pass for PyTorch version 1.8.1 and 1.8.2.

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:

    $ python3.9 -m venv ~/.pytorch-build  # Replace python3.9 with python3.8/python3.6 if building for Python 3.8/3.6
    $ . ~/.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 v1.10.0
    $ 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_BREAKPAD=0 USE_OPENMP=1 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=0 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/.