The current repository contains examples that one could run on the LRZ AI Systems. Instructions on how to create a container to run the examples are mentioned below. Once the container is created, the slurm job scripts provided could be used to run the job on the systems.
The materials of this repo serve as a basic how-to and will also be used as part of the trainings offered.
Follow the steps to extend a container from the NVIDIA NGC Catalog Containers which could be used to run all the examples in this repo.
$ enroot import docker://nvcr.io/nvidia/pytorch:23.06-py3
$ enroot create --name custom_container nvidia+pytorch+23.06-py3.sqsh
$ enroot start custom_container
$ pip install --no-cache-dir lightning
$ pip install --no-cache-dir torchdata portalocker torchvision
$ HOROVOD_GPU_OPERATIONS=NCCL pip install --no-cache-dir horovod
$ exit
$ enroot export --output pytorch+lightning+custom.sqsh custom_container
$ enroot start pytorch+lightning+custom.sqsh
The examples are codes found online adapted to the LRZ AI Systems by Maja Piskac and Ajay Navilarekal Rajgopal (both part of LRZ). Please approach us if we have used one of your works or would like to see one of your works as a part of this repo.