HuggingFace + Ray AIR Integration: A Python developer’s guide to
summary
Shows a working integration between Hugging Face Transformers and Ray AI Runtime: the Ray AIR Hugging Face Trainer runs existing Trainer code distributedly, sharding data with Ray Datasets and using PyTorch DDP/FSDP for gradient sync. Demo runs on a 16‑GPU cluster (4 nodes × 4 GPUs), example config uses 3 workers × 1 GPU, and training finishes returning metrics and a checkpoint. Features include automatic checkpointing and resume, built‑in logging, easy HPO with Ray Tune (retune) and batch prediction via the Hugging Face predictor, all demonstrated end‑to‑end.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:31What we will cover
- 2:13Modern ML state
- 4:22Hugging Face benefits
- 7:36Distributed training necessity
- 10:11Ray AI Runtime
- 13:04Hugging Face + Ray AIR
- 15:11Ray AIR’s trainer
- 16:57Ray Data for scaling
- 18:46Monitoring and checkpointing
- 20:12Converting workflows
- 21:39Hyperparameter tuning
- 22:27Scalable inference
- 23:08Live demo
- 27:21Conclusion and Q&A
related talks
22mRay, a Unified Distributed Framework for the Modern AI Stack
35mOperating a Production-Scale Physical AI Data Platform with Ray
26mA Growing Ecosystem of Scalable ML Libraries on Ray
02h 51mDistributed Python with Ray Hands on with the Ray Core APIs
33mOvercoming Distributed ML Challenges with Ray Train
57mIntroduction to Ray for distributed and machine learning applications in Python
01h 30mDistributed Python with Ray-Hands on with the Ray 2.0 APIs for scaling Python Workloads
45mAccelerate Model Training with an Easy to Use High-Performance...