Creating a PyTorch Environment

Owned by Simon Alexander
Oct 09, 2024
1 min read

Getting Started

The main PyTorch page has an Install PyTorch section.

The most straight forward way to get started with PyTorch is by creating a conda environment. Although it can be pip installed, unless you are comfortable with how Python and Pip installs work across the cluster, we’d recommend conda to create self contained environments, which you can have many of, spanning different versions (and different versions of underlying libraries such as Cuda), that themselves can be built upon with additional packages.

If you have not used conda before, please consider working through the following workshops:

Creating Conda Environments

PyTorch itself comes in two versions, stable and preview.

These versions themselves then can be built with different versions of Cuda and/or CPU only.

Which Cuda version? Which version of Cuda you choose will depend on the type of NVidia GPU device you intend to use, and that device’s compute capability. Bear in mind that older GPUs will not support later versions of Cuda.

[]$ cd /project/<project-name>/software/pytorch/ []$ module purge []$ module load miniconda3/<version> []$ conda create -p pyt_2.4.1_cuda12.4 ()[]$ conda activate /project/<project-name>/software/pytorch/pyt_2.4.1_cuda12.4 ()[]$ conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia ()[]$ conda deaactivate []$