The conference includes four technical workshops led by MINERVA partners.
Generative Multimodal Models on HPC
From text-to-image diffusion to diffusion language models
A practical, 2-hour technical session on the generative models behind today’s multimodal AI,
and on how to train and run them on HPC infrastructure. Participants will run multi-GPU training and inference jobs on Leonardo
through SLURM. They will study the trade-offs between denoising steps, parallel decoding, output quality
and latency, and compare the results against autoregressive baselines on standard benchmarks.
Efficient LLM Training on HPC
Scalability, performance & optimization
A practical, 2-hour technical deep dive into scaling Large Language Model (LLM) training on HPC infrastructure.
Participants will learn how to deploy training jobs on big scale for medium to big size models, identify
hardware and distributed bottlenecks using modern profiling tools, and apply memory and execution
optimizations across multi-GPU and multi-node setups.
Federated Learning at Scale
Federated learning across distributed European HPC infrastructures
This workshop will explore the opportunities and benefits of federated learning at different scales, with particular attention to federated learning across European HPC clusters.
It will provide an overview of current approaches, their limitations, and the challenges that remain for scaling federated learning across distributed HPC infrastructures.
Data Generation for HPC Infrastructures
How agentic data is generated
Agentic models learn from trajectories: multi-turn interactions in which a model plans, calls tools and corrects itself. Generating them at scale is as much a compute problem as a data one. This session shows how the process is organised on HPC infrastructure, from launching generation jobs on Leonardo through SLURM to producing clean, training-ready datasets, and which trade-offs between throughput, quality and diversity matter along the way.
