PyData Trójmiasto #40

40th edition of PyData Trójmiasto coming up!

Where: Łużycka 8B, Gdynia, Tensor Z building.
When: 18th of February at 6PM

Registration: please register with your full first and last name. Don't forget to bring your ID when entering the event venue.
RSVP available until 6pm day before. In any urgent communication please let us know at kontakt@pydata-trojmiasto.pl

Parking: Free parking located along the street next to the building.

Agenda:

18:00 - 18:05 - Event boarding
18:05 - 18:10 - A few words about PyData
18:10 - 18:55 - Traditional Datacenters vs Microsoft AI Superfactory by Sal Rosales
19:00 - 19:45 - vLLM in Practice: Efficient LLM Inference with Paged Attention by Jakub Sochacki
19:45 - Networking & 🍕 🍕 🍕!!!

About Traditional Datacenters vs Microsoft AI Superfactory

This presentation explains why AI workloads require a fundamentally different type of infrastructure than traditional cloud datacenters, using Microsoft’s Atlanta Fairwater AI Superfactory as a concrete example.

Key topics include:
• How traditional datacenters are built to run millions of independent applications, while AI superfactories are designed to run one massive AI workload across hundreds of thousands of GPUs
• Why power density, cooling, and networking become critical constraints for modern AI training
• How Microsoft connects multiple AI datacenters (Atlanta and Wisconsin) into one distributed “virtual supercomputer”
• Why design choices such as liquid cooling, two‑story buildings, and AI‑optimized networks directly affect AI model training speed and scale
• What these infrastructure changes mean practically for AI programmers (faster training, larger models, new workflows)

Participants will leave with a clear mental model of how and why AI infrastructure has evolved, and why future AI capability depends as much on datacenter design as on algorithms and software.

Sal Rosales:

Sal is a Sr. Technical Program Manager based in San Diego, California. He has worked at Microsoft for 14 years and is currently on a team that manages the buildout of datacenters all over the globe. He frequently conducts community outreach workshops to under-represented communities. In his free time, he enjoys working out for half-marathons and triathlons. Participating in these events not only offers a change of pace from work but also helps spur his growth mindset of jumping out of his comfort zone.

About vLLM in Practice: Efficient LLM Inference with Paged Attention:

How do you serve large language models efficiently in production? In this talk, Jakub will introduce vLLM, explain the Paged Attention algorithm behind it, and show why it enables high throughput and better GPU memory utilization. He will also cover key vLLM features, common use cases, and demonstrate how to run LLM inference in practice with a short terminal demo.

Jakub Sochacki:

Jakub is an AI Software Engineer at Intel specializing in optimizing large language model inference performance on AI hardware accelerators. Starting contributor to the vLLM project with a background in machine learning research, including work on LLM hallucinations at Gdańsk University of Technology. Passionate about building and deploying LLM-based systems in practice.