Highlights: Ray Summit 2022
The in-person conferences and community meetups are back. Last week’s Ray Summit 2022, held in-person and partially virtual at the San Francisco Hyatt Regency, was a testament to a reemergence of community gatherings after two years of absence due to the COVID-19 pandemic.
Over 650+ attendees gathered to learn and discover everything about Ray and about scaling AI/ML from keynotes, session speakers, and attendees, crossing oceans and traversing lands from 22 US states and 15 countries.
With over 700+ contributors, a rising GitHub amassed of coveted stars, a growing community, and an increasing cadence of Ray talks at meetups and global conferences, the open-source Ray project and community are establishing a global presence.
Ray training and community Meetup
Typical and expected at open-source conferences, the summit kicked off with three 3 ½ hour training sessions on “Introduction to Ray for distributed applications,” “Model development and deployment with Ray Serve,” and “Introduction to RL and RLlib.” Catered to different levels of expertise, these three classes, taught by seasoned advocates and committers, satiated the knowledge-thirsty Ray beginners and enthusiasts.
Happy hour, followed by a Ray community meetup, is a staple gathering at these conferences, fostering enduring connections. Three community speakers shared their use cases and how they use Ray at scale. Arize’s Aparna Dhinakaran asserted the need for model observability, highlighting integration with Ray; Ikigai’s Jaehyun Sim explained a multi-tenant strategy for long-running Ray jobs; and former assistant professor at Syracuse university discussed how to scale nearest neighbor search with Ray.
Ray summit keynotes assert scaling AI
From the onset, the founders of Anyscale and the creators of Ray always had scaling and making distributed computing easy for AI/ML and Python as their guiding principle. At the summit, CEO and Co-founder Robert Nishihara introduced Ray 2.0 and Ray AIR to make Ray simpler, easier, and performant and to scale simple and common ML workloads.
Co-founder Ion Stoica further asserted that distributed computing is complicated and challenging. Today’s large models require massive amounts of computing power and stitching together bespoke systems is brittle and complex. One solution is capitalizing on blessings of scale. Another is using a unified framework like Ray on Anyscale, providing scalability, manageability, and uniformity of libraries for common ML workloads.
To demonstrate what Ion asserted — the need for a unified toolkit for a typical end-to-end workload like building a recommendation engine for movies — Robert showed how-to develop each machine learning (ML) pipeline stage, step by step, using Ray 2.0 and Ray AIR on the Anyscale platform.
Likewise, other keynotes spoke to similar themes of scaling AI/ML workloads. For example, OpenAI Co-founder and CTO Greg Brockman revealed how Ray helped to train and scale their large language models. That’s a testament of Ray’s potential.
Smitha Shyam, Uber’s head of the ML/AI engineering, declared how Ray helped transform their internal processes and addressed pain points that were previously hard to solve. In particular, scaling Uber ML applications such as UberEats and ETA and Ride ranking.
Adding to the thematics of scale, luminaries Soumith Chintala (co-creator of PyTorch and lead at Meta AI) and Kim Hazelwood (head of Meta AI Research) said scaling AI is a necessity for innovation and sustainability. Dario Gil, senior vice president of IBM research, made a quantum leap linking the use of Ray with the Quantum computing toolkit Qiskit.
And finally, the UC professor of computer science, Anca Dragan, insightfully revealed behavioral connections between robotics and human interactions.
Uses cases and learning from experience
Nothing says more critically or constructively about Ray, its use cases, and its adoption than stories shared by the Ray community: all their pain points, how they resolved it, and what they learned in their Ray journeys. Over 50+ breakout sessions and lightning talks from Uber, Riot Games, Amazon, Shopify, Stitch Fix, Microsoft, Instacart, Spotify, Deepset, IBM, Meta AI, Ant Group, ByteDance, Cruise, Qatar Research/FIFA2022, Lyft, Netflix, Dow, Minds.ai, Predibase, Intel, etc. spoke volumes about Ray. Not only do these notable companies use Ray at scale, but also how Ray unleashed additional use cases.
What next?
All keynotes, sessions, lightning, and meetup talks are available on-demand. To help you navigate, we have guides for you: Reinforcement learning sessions at Ray Summit and 7 must-attend Ray Summit sessions: RL-powered traffic control, infra-less ML, and more. For an ML developer interested in specific Ray 2.0 features, I suggest the following talks:
- Introduction Ray AI Runtime
- State of Ray Serve in 2.0
- Shuffling 100TB with Ray Datasets
- Deep dive into data ingest with AIR + Datasets
- Ray Observability: Present & future
- Many others in Ray Deep Dives track
Thank you if you attended in person or online, spoke at the summit, or sponsored the summit. Here is a quick clip from day 1, and what the Ray community is saying about their experience at the summit.
What is the Ray community saying about Ray Summit 2022
Once again, kudos to the program and organizing committee and the Eventi productions. See you again next year!
