Keynote Q&A
summary
Ray splits into a core plus libraries model: use the libraries for common ML workflows, but implement novel distributed algorithms or custom hardware support on the Ray core to scale further. Ray provides a single-system deployment surface that unifies data processing, training, and serving and preserves clear fault-tolerance and consistency semantics. Remote actors are Python classes invoked with object references/futures to avoid large copies. Small domain-specific models (even ~7B) fine-tuned for APIs like Gorilla can beat larger general models for targeted tasks.
AI-written summary · about
chapters (6)
- 0:00Q&A
- 0:24Ray vs core for AI
- 2:55Ray’s AI deployment
- 5:08Ray vs Dapr actors
- 7:59Smaller AI models future
- 10:10Closing
related talks
24mRay: A System for High-performance, Distributed ML Applications
01h 31mIntroduction to Ray for distributed and machine learning applications in Python
15mRay: A Distributed Compute Engine for AI
14mKeynote: The Future of Ray
15mRay: A System for Distributed Applications
26mRay: A System for High-performance, Distributed Python Applications
42mRay: A distributed system for emerging AI applications
26mRay: A General Purpose Serverless Substrate?