Beginner's guide to EIDF: navigating the GPU Service and Kubernetes

Operated in Edinburgh by EPCC, the Edinburgh International Data Facility (EIDF) brings together the things beginners usually struggle to stitch together on their own: access permissions, programming environments, GPU capabilities, and storage options that scale from quick experiments to serious research workflows.

The tutorial Beginner's Guide to EIDF: Navigating the GPU Service and Kubernetes provides a walkthrough for people who want to start with something approachable (lightweight LLM inference) and then build confidence towards training and fine-tuning. You don't begin by wrestling with drivers, CUDA installs, or machine setup. Instead, you follow a clear path: get access to EIDF through EPCC SAFE and the EIDF Portal, pick your preferred working style (SSH terminal or browser-based VDI), and understand the concepts underpinning the GPU Service. You run your work as containers on Kubernetes, inside your project’s own namespace.

This tutorial is a beginner-friendly runway: start small with inference, adopt good platform habits early, and build toward training-ready workflows on infrastructure that’s designed to support you as your ambitions grow.

Learning habits that scale

From there, the tutorial quickly delivers an early win: launching a GPU-backed PyTorch environment from a popular container image, then stepping inside the running container to confirm that GPUs are visible and usable. That milestone matters because it turns "I have access" into "I can run GPU code" in a single session — plus it sets you up with habits that scale (explicit resource requests, named jobs, and a clean submit/monitor/stop cycle).

The fun part!

Once you’ve got a pod running, the tutorial pivots into the fun part: LLM inference with a small, downloadable model using the Hugging Face ecosystem. It’s deliberately lightweight — install a few Python packages, load a model, ask questions against a snippet of context — so you can focus on the workflow rather than the weight of the model. And it uses this moment to teach two lessons that matter as you progress: containers are disposable (so you’ll eventually want custom images or persistent storage), and LLMs can be confidently wrong (with maths errors being a classic example).

Creating a research workflow

Finally, the tutorial shows you how to build up from "command-line poking" to a comfortable research workflow: running JupyterLab on the GPU Service and reaching it securely via port-forwarding, then adding persistent storage so notebooks, datasets, and checkpoints survive beyond the lifetime of a single job. That’s the bridge from tinkering to iteration and it’s exactly what you need when you move from inference into training and fine-tuning, where reproducibility, data management, and checkpoint persistence stop being “nice to have” and become essential.

https://getting-started-with-the-gpu-service-and-llms-fbe560.pages.eidf.ac.uk/

UK AI Factory Antenna and EIDF

EPCC has won EU and UK Government funding to establish and operate the UK AI Factory Antenna (UKAIFA) and is part of the broader push to make advanced AI capability accessible and usable across the UK research community.

• UKAIFA (Antenna) will help organisations (often an industry partner) shape a viable AI adoption path: what they want to do, what success looks like, what constraints exist, and what compute/service model they need.
• EPCC will provide the technical bridging: translating that plan into an implementable approach (containers, data movement, security posture, scaling expectations), and helping users navigate the platform effectively.
• EIDF is one of the delivery platforms where that work will run: projects will be created/joined, budgets/charging routes will be agreed, accounts will be issued, and GPU + storage services will be consumed to run inference, fine-tuning, and potentially larger training workflows.

Edinburgh International Data Facility (EIDF)

Tutorial: Beginner's Guide to EIDF: Navigating the GPU Service and Kubernetes

By: Dr Amy Krause |