PyTorch on HPC

We have PyTorch (1.4 & 1.7) installations as part of Miniconda module on the HPC. You can clone the existing centralized environment to your local environment and add other required libraries or packages on top of it after activating it. To find more details of the Miniconda module on the HPC, click here.

Note

If you have never used Conda, we recommend you to use the HPC Miniconda. You can find the steps to set up the HPC Miniconda by clicking here.

How to clone the PyTorch environment

1. If you are using the HPC Miniconda

   #conda create -n <name of the new env> --clone <existing env>
   #example:
   conda create -n pytorch --clone pytorch-1.11.0
   #You can also have the environment in a custom location
   conda create -p /scratch/wz22/conda-envs/pytroch
   

2. If you are using your own conda package

   #conda create -n <name of the new env> --clone <path to existing env>
   #example:
   conda create -n pytorch --clone /share/apps/NYUAD5/miniconda/3-4.11.0/envs/pytorch-1.11.0
   

Submitting Job Scripts

The conda environment might not get activated when submitting a Job script since the slurm doesn’t source the bashrc file. Hence, in order to go about this, you can include the following line in your job submission script before activating the required environment.

source /share/apps/NYUAD5/miniconda/3-4.11.0/bin/activate

A sample job submission script is shown below:

#!/bin/bash
#SBATCH -c 10
#SBATCH -t 48:00:00
#SBATCH -p nvidia
#SBATCH --gres=gpu:1
#Other SBATCH commands go here

#Activating conda
source /share/apps/NYUAD5/miniconda/3-4.11.0/bin/activate
conda activate pytorch

#Your appication commands go here
python abc.py

Note

These installations have preinstalled cuda and cudnn libraries as well and hence there is no need to load cuda modules explicitly