Conda environments

TL;DR

• Conda handles installing PyTorch with CUDA

environment.yaml

This is a minimal conda environment config environment.yaml which will install PyTorch with CUDA 11.8

name: ENV_NAME
channels:
  - nvidia
  - pytorch
  - conda-forge
dependencies:
  - pytorch=2.0
  - pytorch-cuda=11.8
  - torchvision=0.15
  - wandb=0.15
  - hydra-core=1.3
  - hydra-submitit-launcher=1.2

Create environment

On a cluster we usually need to load a module for minoconda with

module load miniconda

It varies from cluster to cluster. As conda is actually quite slow I prefer to use mamba. We can create an environment from this environment config environment.yaml with

mamba env create -f environment.yaml

If we add to environment.yaml we can easily update it with

mamba env update -f environment.yaml