Keynote – Ion Stoica
summary
Claims that Ray consolidates streaming, training, tuning and serving into one system and replaces stitching disparate frameworks, with a minimalist six-call API. Core features are tasks and actors for stateless and stateful work and a distributed object store that avoids extra data hops. Cites 35x compute growth every 18 months versus GPU memory 1.7x, models up to 1.75 trillion parameters, 20,000+ GitHub stars, and deployments on 200,000 cores for large events.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 0:29Key trends in distributed AI
- 1:55ML demands vs hardware
- 3:43AI applications in industries
- 5:56Challenges in ML systems
- 7:22Ray for distributed ML
- 9:13Ray’s minimalist API
- 10:39Tasks actors and futures in Ray
- 13:28Ray’s in-memory store
- 16:36Ray’s industry adoption
- 16:56Ray use cases at scale
- 19:06Summary
related talks
17mKeynote Analysis
02h 16mKeynote Day 2 - Full
35mHow Bytedance Builds Large-Scale Data Processing Pipelines for Multimodal Models with Ray
20mKeynote Review
26mData Processing on Ray
32mAnyscale's Ray Data: Revolutionizing Batch Inference
31mFrom OOMs to 2x Throughput: Scaling Multimodal Data Curation on Ray
30mHow to Scale Unstructured Data Processing with Ray Data