Running on LUMI

LUMI is a EuroHPC supercomputer hosted at CSC’s data center in Kajaani, Finland. It is owned by the European Commission’s EuroHPC Joint Undertaking and operated by a consortium of 11 European nations. With a sustained computing power of 380 petaflops, LUMI ranks among the fastest supercomputers in the world (9th on the Top500 list, November 2025). The system runs on 100% renewable energy, and its waste heat is used to warm the city of Kajaani. All models in this repository (MNIST, USPS, SVHN) were trained on LUMI-G using AMD MI250x GPUs.

This page explains how to run training and evaluation on LUMI using the scripts provided in the lumi/ directory. For a broader introduction to running AI workloads on LUMI, see the Getting Started with AI on LUMI workshop repository.

Environment setup

For a detailed guide on setting up your environment on LUMI, see the LUMI AI Guide — Setting up the environment.

The scripts in this repository use a Singularity container that provides PyTorch with ROCm support for AMD GPUs. The software environment is loaded inside each Slurm script automatically — no manual module load or pip install steps are needed.

Training

The repository includes a ready-made Slurm script at lumi/train.sh. Submit it from the repository root:

# Train the default dataset (MNIST)
sbatch lumi/train.sh

# Train a different dataset by overriding the job name
sbatch -J svhn lumi/train.sh
sbatch -J usps lumi/train.sh

The job name (-J) selects the experiment: it loads configs/<name>.cfg for hyperparameters and names the log files lumi/logs/<name>_<job-id>.{out,err}.

To add a new dataset, create a matching configs/<name>.cfg file alongside the existing ones.

Evaluation

A separate Slurm script at lumi/evaluation.sh evaluates a trained checkpoint on the test set. Submit it from the repository root:

# Evaluate the default dataset (MNIST)
sbatch lumi/evaluation.sh

# Evaluate a specific dataset
sbatch -J eval-svhn lumi/evaluation.sh
sbatch -J eval-usps lumi/evaluation.sh

The script derives the dataset name and checkpoint path from the job name (eval-<dataset> → src/weights/<dataset>.pth). If the checkpoint is not found, it exits with an error listing the available checkpoints.

Note: Inference speed reported on LUMI may appear slower than on a local machine. This is expected — HPC environments introduce overhead from container startup, Slurm job scheduling, and parallel filesystem I/O that is not present in local runs.

Monitoring jobs

# Check the status of your running jobs
squeue --me

# View the full log output of a job
cat lumi/logs/<job-name>_<job-id>.out

# Follow the log output in real time (while the job is running)
tail -f lumi/logs/<job-name>_<job-id>.out

Tips

  • Both scripts can also be submitted from inside the lumi/ subdirectory:

    cd lumi && sbatch train.sh
cd lumi && sbatch -J eval-svhn evaluation.sh

  • The --account in the scripts is set to project_465002757. Update this if you are running under a different project allocation. Check your allocations with csc-projects.

  • Training time is capped at 30 minutes and evaluation at 10 minutes in the default scripts. Adjust #SBATCH --time if needed.