A Growing Ecosystem of Scalable ML Libraries on Ray
summary
Covers concrete integrations: Horovod is checked into its repo to use Ray for process placement and autoscaling, HuggingFace’s RAG moves retrieval into Ray actors and gains up to 2x faster document retrieval in multi‑GPU fine tuning, and higher‑level trainers like PyTorch Lightning and spaCy use Ray for placement groups, autoscaler‑driven elastic training and cluster launch. Examples include actor-based retrieval, Horovod + RayExecutor code flow, and support for fairscale sharded training across nodes.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:30Necessitated by 2 Trends
- 3:00Specialized Hardware is not enough
- 4:27Cutting edge approaches require ad-hoc distributed computation
- 6:00Elastic Training and failure handling can be complex
- 8:26Three key benefits
- 10:42Ray Placement Groups
- 11:41Ray Autoscaler
- 12:50Ray unifies the distributed ML Ecosystem
- 16:57Horovod on Ray Adoption • Integrated as a backend in the Horovod repo
- 18:26Document Retrieval with Torch Distributed
- 20:35Huggingface Transformers
- 21:10Pytorch Lightning
- 21:56Ray Lightning Library
- 23:27Ray Lightning Architecture
related talks
27mDistributed Deep Learning with Horovod on Ray
24mRay: A System for High-performance, Distributed ML Applications
28mScaling Multimodal Data and Reinforcement Learning with Ray
29mScaling ML Pipelines for Autonomous Driving with Ray
22mRay, a Unified Distributed Framework for the Modern AI Stack
01h 31mIntroduction to Ray for distributed and machine learning applications in Python
35mOperating a Production-Scale Physical AI Data Platform with Ray
33mOvercoming Distributed ML Challenges with Ray Train