Moving your AI training jobs to LUMI workshop - CWI, Amsterdam, May 27-28, 2025¶

Course organisation¶

Location: CWI, Science Park 123, 1098 XG Amsterdam, The Netherlands
The public transportation is run by GVB.
- Whatever type of ticket you have, you always need to check in and check out.
  
  It is now perfectly possible to do so with your bank card, but of course you should use the same one to check in and check out.
  
  If you use your bank card, you can use their app Gappie to follow up your use. The app also provides routing, and you can by regular single or multi-day tickets in it that will then work with a QR code.
- Gappie app for iOS and Gappie app for Android
- Multi-day tickets are interesting if you take public transportation a lot. You can buy them in the Gappie app, but opening the app and scanning the QR code is a slow process compared to tapping a credit card on the reader. If all travel you need is from a hotel in the centre to the venue and back, it isn't worth it.
Amsterdam is more than canals and the red light district. You don't have to drink only water and milk and there is more food than "bitterballen" from the FEBO.
Preliminary schedule

Setting up for the exercises¶

During the course¶

If you have an active project on LUMI, you should be able to make the exercises in that project. To reduce the waiting time during the workshop, use the SLURM reservations we provide (see above).

You can find all exercises on our AI workshop GitHub page

After the termination of the course project¶

Setting up for the exercises is a bit more elaborate now.

The exercises as they were during the course are available as the tag ai-20250527 in the GitHub repository. Whereas the repository could simply be cloned during the course, now you have to either:

Download the content of the repository as a tar file or bzip2-compressed tar file or from the GitHub release where you have a choice of formats,

or clone the repository and then check out the tag ai-20250527:

git clone https://github.com/Lumi-supercomputer/Getting_Started_with_AI_workshop.git
cd Getting_Started_with_AI_workshop
git checkout ai-20250527

Or you can download them as a .tar.gz file or ZIP-file.

Note also that any reference to a reservation in Slurm has to be removed.

The exercises were thoroughly tested at the time of the course. LUMI is an evolving supercomputer though, so it is expected that some exercises may fail over time, and modules that need to be loaded, will also change as at every update we have to drop some versions of the LUMI module as the programming environment is no longer functional. Likewise it is expected that at some point the ROCm driver on the system may become incompatible with the ROCm versions used in the containers for the course.

Course materials¶

Note: Some links in the table below will remain invalid until after the course when all materials are uploaded.

Presentation	Slides	recording
Welcome and course introduction	/	video
Introduction to LUMI	slides	video
Using the LUMI web-interface	slides	video
Hands-on: Run a simple PyTorch example notebook	/	video
Your first AI training job on LUMI	slides	video
Hands-on: Run a simple single-GPU PyTorch AI training job	/	video
Understanding GPU activity & checking jobs	slides	video
Hands-on: Checking GPU usage interactively using rocm-smi	/	video
Running containers on LUMI	slides	video
Hands-on: Pull and run a container	/	video
Building containers from Conda/pip environments	slides	video
Hands-on: Creating a conda environment file and building a container using cotainr	/	video
Extending containers with virtual environments for faster testing	slides	video
Scaling AI training to multiple GPUs	slides	video
Hands-on: Converting the PyTorch single GPU AI training job to use all GPUs in a single node via DDP	/	video
Extreme scale AI	slides	video
Demo/Hands-on: Using multiple nodes	/	video
Loading training data on LUMI	slides	video
Coupling machine learning with HPC simulation	slides	video
Hands-on: Advancing your project and general Q&A	/	video

Web links¶

LUMI documentation