Learn how to run AI coding agents safely with the LUMI supercomputer, using open models on Aitta, the OpenCode agent and the LUMI MCP server.

Learning goals

• After the training, you will be able to:
• Use open-weight LLMs on Aitta, the LUMI AI Factory inference platform, through the web interface and the OpenAI-compatible API,
• Run the OpenCode coding agent on LUMI in the ready-made LUMI AI Factory container, connected to Aitta and the LUMI MCP server,
• Understand the difference between an LLM, a harness and a coding agent, and what each of them is responsible for,
• Use an agent to write and fix Slurm batch jobs, port code to LUMI’s AMD GPUs and find performance problems, while you stay in charge of what runs on the shared system,
• Recognise the risks of AI agents, such as hallucinations and prompt injection, and know how to work around them.
                 

Target audience

Researchers, students and professionals from companies and the public sector who want to get to know the LLM and coding agent services of CSC and the LUMI AI Factory. You’ll learn how to use our inference platform Aitta, and how to run, configure and safely use a coding agent on LUMI. Along the way, we’ll look at what risks agents bring and what each part of the system is responsible for.
               
                           

Prerequisites

Interest in LLMs and coding agents, curiosity about how they work and how to run them safely.
Nice to have: basic command-line skills, Python, and experience with Slurm or PyTorch for the more advanced exercises. We’ll go through every command together, so none of these are required.

Agenda

• LUMI, the LUMI AI Factory, and a quick intro to HPC and Slurm
• How LLMs and coding agents work: what a bare LLM can do, what a system prompt adds (for example, the current date and time), and what the harness is responsible for
• Aitta: LUMI AI Factory’s inference platform
• OpenCode on LUMI, and how to use it safely
• The LUMI MCP server, and what’s next: FirecREST
• Demo: from logging in to your first agent session
• Hands-on exercises, worked through together on screen: prompt injection, hallucinations with and without the LUMI documentation, moving code from a laptop to LUMI, and finding out why training is slow

Wrap-up and discussion
               

Schedule

Presentation + Demo 09:00 – 10:00
Coffee break 10:00 – 10:15
Hands on exercises 10:15 – 12:00

Practicalities

A laptop and curiosity