Scaling Multimodal Data and Reinforcement Learning with Ray
summary
Claims that a Kubernetes + Ray + PyTorch + vLLM stack handles heterogeneous AI workloads, with Ray providing finer-grain scheduling, failure recovery, and streaming CPU→GPU pipelines via Ray Data and GPU RDMA object transfer. Reports weekly downloads rising from 1.5M to 7M, Amazon cuts latency by ~90%, Geotab gets 43x video throughput and 40% fewer GPUs, Pronovo sees 5x faster development and 40% cost reduction, and multiple RL and agent frameworks run production workloads on Ray.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker background
- 0:52Evolution of computing stacks to AI era
- 3:17Multimodal data and RL pipeline changes
- 5:24Challenges in AI infrastructure stack
- 7:28Transition to PAIR stack concept
- 8:32Ray Data for multimodal processing
- 13:05Ray for reinforcement learning
- 17:02Building AI infrastructure options
- 19:13Ray adoption and customer success
- 22:23Three takeaways summary
- 23:45Resources and Q&A start
related talks
29mScaling ML Pipelines for Autonomous Driving with Ray
35mOperating a Production-Scale Physical AI Data Platform with Ray
24mRay: A System for High-performance, Distributed ML Applications
23mAn Open Source Post-Training Stack: Kubernetes + Ray + PyTorch + vLLM
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
22mLarge Scale Data Loading and Data Preprocessing with Ray
22mRay, a Unified Distributed Framework for the Modern AI Stack
16mKeynote: An Open-Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLM