MLOpstober Edition

Agenda
*******
18:00 - 18:15; Reception
18:15 - 18:20; A welcome note from organizers & sponsor
18:20 - 18:50; Talk 1: Titus von Köller welcomes us to the World of QLoRa: Fine-tune Your Personal AI on Consumer-grade Hardware
18:50 - 19:00; Q&A
19:00 - 19:30; Talk 2: Sarah Holschneider from Eloquest: Expectation Management in AI Projects
19:30 - 19:40; Q&A
19:40 - 19:45; Wrap-up and Announcements
19:45 - 21:00 Socializing with food & drinks (from T-SYSTEMS)
21:00 End of the Event
******
Talk 1: Titus von Köller welcomes us to the World of QLoRa: Fine-tune Your Personal AI on Consumer-grade Hardware

Welcome to the world of QLoRA, where Quantized models meet Low Rank Adapters (LoRA). LoRA, small yet powerful adapters (fundamentally just small NNs), are integrated into larger neural networks, enabling efficient fine-tuning without modifying the base model.
Drawing from a 2023 academic paper and the personal experience contributing `bitsandbytes` Python library, this talk will try to demystify how QLoRA achieves exceptional performance with limited resources and seek to inspire others to explore use-cases, like exploring characters for creative writing, complying with GDPR or researching AI alignment and gender bias.
The magic of QLoRA lies in its ability to drastically reduce memory requirements—fine-tuning an advanced 65B parameter model, like Llama 2 or Falcon, needs less than 48GB of GPU memory (reducing from >780GB), without compromising performance. This efficiency brings advanced AI applications, comparable to ChatGPT 3.5, within reach of consumer-grade hardware, fostering innovation across diverse fields. Surprisingly, this is often already effective with 10s to 100s of training examples.
About the Presenter - Titus:
Two months ago, Titus embarked on a four-month non-profit workation, dedicating his full-time efforts to contribute to the `bitsandbytes` open source library, led by expert deep learning researcher Tim Dettmers. Coming from a background in MLOps, he's venturing into deep learning, a field new to Titus, by collaborating with Tim to help in making deep learning more accessible. This endeavor reflects Titus' commitment to learning and making a meaningful impact on society.
Titus' journey in computing began with a BSc in Psychology, exploring Python due to an interest in neuroscience, which eventually led him to pursue an MSc in Computer Science. In 2017, he founded the PyData Zurich meetup. He has worked on a diverse range of projects, including agent-based modeling, infra-as-code, DevSecOps, Kubernetes, operationalizing ML models, and latest refining developer experience in Continuous Training pipelines.

Talk 2: Sarah Holschneider from Eloquest: Expectation Management in AI Projects - How can I manage expectations when leading an AI project? What can I expect from my internal or external partners? What pitfalls might await me?