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 v1.10.0 and it includes Pytorch 2.0.1.
- CPU version optimized for Power10
- pytorch-cpu 2.0.1 of Python 3.11
- Cuda versions
- For Cuda 12.2
- pytorch 2.0.1 of Python 3.11
- pytorch 2.0.1 of Python 3.10
- pytorch 2.0.1 of Python 3.9
- Cuda 11.8
- pytorch 2.0.1 of Python 3.11
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=2.0.1
Once the installation is success, validate the installation by running the following command which lists the package pytorch which was pulled from rocketce channel
GPU Variant
To install pytorch built for GPU support run the following command:
- conda install -c rocketce pytorch=2.0.1=cuda12.2* - For CUDA 12.2 version
- conda install -c rocketce pytorch=2.0.1=cuda11.8* - For CUDA 11.8 version
Once the installation is success, validate the installation by running the following command which lists the package pytorch which was pulled from rocketce channel
Note: For GPU Cuda 12.2, tensorflow and pytorch will show conflict while installing together. Individually they will install and work without any issues.
Pytorch conda packages
Pytorch package from RocketCE also includes the following conda packages
- pytorch-cpu / pytorch
- pytorch-base
- Sentencepiece
- torchtext
- torchvision-base
The package-base packages come with CPU variants, and include CPU in the build string.
More information
This conda CPU package is using the Power 10 MMA capabilities. More information is available in Matrix-multiply assist (MMA) capabilities.
------------------------------
Rajesh Nukala
Rocket Internal - All Brands
------------------------------