AI Training: From Pytorch to GPU Clusters
AI Training: From Pytorch to GPU Clusters
We're Homebrew, an AI Research Lab. We're the creators and lead maintainers of a few AI tools and models:
Jan: Personal AI (>1.3mn downloads)
Cortex: Local AI engine
llama3-s: Llama3 + Native Speech ability
In this talk, we'll share how we are training llama3-s, our fusion model with native audio comprehension.
We'll walk through our training and data methodology and do a few live demos. We'll also address the challenges we faced at different layers of the stack: from software to the underlying hardware.
Finally, this will include a hands-on hardware demonstration of a small training workstation.
What we'll cover
How does AI training work at a hardware level?
Training: Why did GPT-4 cost $100mn to train?
Hardware: Why did Elon Musk spend $4bn to build a 100,000 GPU cluster?
Datacenters: Why is there a shortage of AI-ready datacenters?
Event Details:
📅 30 July 2024, Tuesday
🕒 6.30pm-8pm
📍 National Library Board, Level 7 - Launch Programme Room
Note: no food (it's a library)
Agenda:
6.30pm: Doors open & Networking
7.00pm: From Pytorch to GPU Clusters
8.00pm: Optional - grab dinner!
Speakers
Nicole Zhu, Co-founder, Homebrew
Nicole Zhu is the co-founder of Homebrew. Previously, Nicole has worked at Google, GoJek, IDEO and studied Computer Science at Stanford.
Gabrielle Ong, Core Team, Homebrew
Gabrielle is an engineer at Homebrew. Previously, she was a solutions architect at AWS, a machine learning researcher at the University of Toronto, and a software engineer at TradeGecko.