Skip to main content

Overview 

Torchaudio is an open-source library to help you develop and apply PyTorch to the audio domain.

RocketCE includes TorchAudio package of version 0.13.0 which is suitable for PyTorch version 1.13.0. This package is available in channel rocketce of anaconda.org (https://anaconda.org/rocketce/repo).  

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


CPU Support
 

RocketCE includes a version of TorchAudio built without GPU support. This inclusion allows for training and inferencing to be done on Power 9 and Power 10 systems that do not have GPUs, or on systems where you want to train and inference without using the GPUs.

 

GPU Support 

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

 

Installation instructions 

TorchAudio has a dependency of Pytorch. Based on local conda configuration setting, installation of TorchAudio may pull and install Pytorch from different channel other than rocketce. So, it is recommended to use the below installation command with version and build id.

CPU Only Variant 

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

  • conda install -c rocketce torchaudio=0.13.0=cpu* pytorch=1.13.0=cpu*

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

  • Conda list torchaudio=0.13.0

GPU Variant 

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

  • For CUDA 11.2 version
    • conda install -c rocketce torchaudio=0.13.0=cuda11.2* pytorch=1.10.2=cuda11.2*
  • For CUDA 11.4 version
    • conda install -c rocketce torchaudio=0.13.0=cuda11.4* pytorch=1.13.0=cuda11.4*

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

  • For CUDA 11.2 version
    • Conda list torchaudio=0.13.0=cuda11.2*
  • For CUDA 11.4 version
    • Conda list torchaudio=0.13.0=cuda11.4*

Dependencies

Following are the corresponding TorchAudio and supported Pytorch and Python version mapping

TorchAudio

Pytorch

Python

CPU

0.13.0

1.13.0

>=3.8, <=3.10

CUDA 11.2

0.13.0

1.10.2

>=3.8, <=3.10

CUDA 11.4

0.13.0

1.13.0

>=3.8, <=3.10


TorchAudio conda packages
 

Following packages are available in rocketce channel and gets installed with TorchAudio package 

  1. Sox
  2. Mad
  3. Speex
  4. lame 

 

Following packages are available in rocketce channel and requires to install to support some of the TorchAudio operations
 
       1. pysoundfile



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