Transforming Grab’s AI Platform with Ray
summary
Grab consolidates offline training, simulator parameter search, and RL on one Ray layer, using Ray Train, Data, Tune, and Core plus Optuna to automate searches and resource placement. Concrete results: identical training code runs on 140M to 5.2B models across 8 A100s, 2x4 sharded ATNGs, or single‑GPU clusters; constrained search runs with Optuna via Ray Tune automating trials and guardrails; simulator RL runs 30 parallel simulators producing ~600M trajectories (~100 GB) streamed into trainers, training loops finish in about 3 hours with automatic restarts and measurable cost/throughput tradeoffs.
AI-written summary · about
chapters (6)
- 0:00Introduction to Grab’s AI platform
- 2:20Offline training with Ray
- 11:10Simulation workflow with Ray
- 17:08Reinforcement learning in simulator
- 24:53Future of integrated AI platform
- 27:41Q&A session
related talks
22mRay, a Unified Distributed Framework for the Modern AI Stack
21mHow Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery
28mScaling Multimodal Data and Reinforcement Learning with Ray
29mScaling ML Pipelines for Autonomous Driving with Ray
44mPanel: ML Platform on Ray
28mScaling AI & Machine Learning Workloads With Ray on AWS, Kubernetes, & BERT
38mNetflix's Ray Platform: From Deep Learning to GenAI
29mHow Roblox Scaled Machine Learning by Leveraging Ray for Efficient Batch Inference