RocketCE for Power

 View Only

Releasing RocketCE Pytorch v1.10.1 with OpenCE v1.5.1

  • 1.  Releasing RocketCE Pytorch v1.10.1 with OpenCE v1.5.1

    ROCKETEER
    Posted 09-07-2022 06:51

    Overview 

    RocketCE Pytorch includes support for IBM's Distributed Deep Learning (DDL).  

    This version of Pytorch will have the support of matrix-multiply assist (MMA) which was introduced in Power 10. 

    This release of RocketCE is based on Release OpenCE v1.5.1 and it includes Pytorch 1.10.1. 

    These packages are supported for both Power 9 and Power 10 Systems of IBM with linux-ppc64le. 

     

     CPU Support 

    The CPU-only variant of Pytorch is built without GPU and CUDA support. It has a smaller installation size, and omits features that would require a GPU. This can be used in both Power 9 and Power 10 systems.

     

    GPU Support 

    This version of Pytorch will support GPU based operations. This can be used in Power 9 systems.

      

    Installation instructions 

    CPU Variant 

    To install pytorch built for CPU support run the following command: 

    • conda install -c rocketce pytorch-cpu=1.10.1  

     Once the installation is success, validate the installation by running the following command which lists the package pytorch which was pulled from rocketce channel  

    • Conda list pytorch-cpu

    GPU Variant 

    To install pytorch built for GPU support run the following command: 

    • conda install -c rocketce pytorch=1.10.1  

     Once the installation is success, validate the installation by running the following command which lists the package pytorch which was pulled from rocketce channel  

    • Conda list pytorch

    Pytorch conda packages 

    Pytorch package from RocketCE also includes the following conda packages 

    1. pytorch-cpu / pytorch
    2. pytorch-base 
    3. Sentencepiece 
    4. torchtext 
    5. torchvision-base 

    The package-base packages come with CPU variants, and include CPU in the build string. 

      

    More information
    This conda package is using the Power 10 MMA capabilities. More information is available in Matrix-multiply assist (MMA) capabilities.



    ------------------------------
    Rajesh Nukala
    Rocket Internal - All Brands
    ------------------------------