Distributed AI Training and Inference Isambard-AI Workshop - Bristol Data Week 2026¶
Abstract¶
Modern AI models have grown too large to fit on a single GPU making it essential to spread work across multiple powerful datacenter-grade GPUs. This hands-on workshop introduces participants to running AI workloads at scale on Isambard-AI, one of the UK's most powerful supercomputers.
Firstly, using PyTorch, we explore how to train AI models across multiple nodes simultaneously. Participants will run a real training benchmark and observe how different communication strategies affect performance.
Secondly, we turn to inference with vLLM — the process of getting predictions out of a trained model. We look at how to serve a Large Language Model (LLM) across multiple GPUs, and measure how quickly the model can respond to requests.
No prior supercomputing experience is required. Participants should have some familiarity with training AI models in Python to get the most out of the workshop.
Registration¶
The Bristol Centre for Supercomputing (BriCS) is running a workshop on distributed AI training and inference on Isambard-AI as part of Bristol Data Week at The Bill Brown Design Suite, Queen's Building, University of Bristol on Tuesday, 2nd June 2026 from 13:00-16:00.
The workshop is open to participants from the following:
- UK research organisations (e.g. universities and research institutes .ac.uk)
- UK industry (UK registered business of any size with a Companies House registration number)
- UK Government
- NHS
- UK registered charities
In order to take part in the workshop and gain access to the system, you must:
- Pre-register for the workshop by filling in the following form by 5PM on Thursday the 28th May 2026
- Bring your own laptop
Warning
If you have not pre-registered you will not be able to access the system on the day.
You must register with your institutional or professional email address. Strong compliance rules mean that invitations to personal email addresses (e.g. Gmail, Yahoo, Outlook) are not allowed.
Required pre-workshop steps¶
If you have completed the above successfully, you will receive an email from [email protected] with the subject "Invitation to the Bristol Data Week 2026 Workshop project".
After registration and receiving the invite, you will be contacted with some steps to do before the workshop. You should follow steps 1-4 after you receive an invitation email.
If you have an account with a research institute supported by MyAccessID (you can check this at https://mms.myaccessid.org/fed-apps/profile/) then you should use this option. When logging in to any Isambard services, you should always choose "University Login (MyAccessID)". If not, check the "Non-research affiliation" tab.
If your institution or company does not support EduGAIN/MyAccessID as a federated login solution, we will have to create you an account in our "identity provider of last resort" (hosted in AWS).
You will receive an email from [email protected] with the subject "Invitation to join AWS IAM Identity Center" asking you to finalise creating an account, using the same email address that you signed up for the workshop.
Once that account is set up and working, from now on when working with Isambard you should always select the "Other Login (IdP of last resort)" option on the login screen.
Warning
You must complete the required pre-workshop steps and set your UNIX username in advance of the workshop.
Access will not be allowed for online attendees. You must attend the workshop in person. You should promptly inform us if you can no longer attend the workshop by emailing us.
This workshop has been allocated a finite amount of compute time to be shared between all participants. Users must not run long-running or large-scale jobs before the workshop and should use the resources as instructed during the workshop.
Workshop Material¶
The workshop will include hands-on sessions delivered by BriCS on the Distributed PyTorch Training tutorial and the Distributed vLLM Inference tutorial.