Ray Train
ray.io
- Free plan
- Windows
- macOS
- Linux
- Self-hosted
What Ray Train is
Ray Train distributes model-training compute among worker processes in a Ray cluster. A training function, workers, a CPU or GPU resource configuration, and a Trainer run distributed jobs. It supports workloads for generative AI foundation models, time-series models, and traditional machine learning such as XGBoost. Integrations include PyTorch, PyTorch Lightning, Hugging Face Transformers, XGBoost, JAX, DeepSpeed, TensorFlow and Keras, LightGBM, and Horovod. Ray Train works with Ray Data for streaming input and preprocessing, while framework-native utilities include PyTorch Dataset and Hugging Face Dataset. Ray Data can spread heavy preprocessing across CPU nodes, and Ray Train can split data among workers during execution. Guides cover metric monitoring, logging, and experiment tracking. Ray describes scaling from a laptop to thousands of GPUs, including heterogeneous CPUs and GPUs that scale independently. It supports Linux, macOS, Windows, and self-hosted deployment, and is listed as free. Built-in token authentication starts in version 2.52.0, but jobs are not isolated and developers lack cluster access controls.
Who it suits
Ray Train suits Python teams running distributed training across CPU or GPU resources. It fits users working with supported machine-learning frameworks and Ray clusters.
Good to know, and what to watch
- Distributed training across Ray workers
- Integrates with many training frameworks
- Supports streaming data and distributed preprocessing
- Handles heterogeneous CPU and GPU resources
- Includes monitoring and experiment tracking guidance
- No job isolation within a cluster
- No developer access controls within a cluster
- Security guidance recommends controlled networks
- Requires separate clusters for workload isolation
Plans
Every plan and priceRay Train
Free plan
The verdict
Ray Train offers broad framework integration and distributed resource control for model training. Plan cluster security carefully because jobs and developer access are not isolated.
What the maker publishes
- Data integration
- Ray Train integrates with Ray Data for streaming data loading and preprocessing, and also supports framework-native data utilities such as PyTorch Dataset and Hugging Face Dataset. source
- Experiment tracking
- Ray Train has an experiment tracking user guide. source
- Framework integrations
- Ray Train integrates with PyTorch, PyTorch Lightning, Hugging Face Transformers, XGBoost, JAX, DeepSpeed, TensorFlow and Keras, LightGBM, and Horovod. source
- Intended users
- Ray’s security documentation describes Ray developers running local single-node clusters or remote multi-node clusters on infrastructure provided by platform providers. source
- Monitoring
- Ray Train provides user guides for monitoring and logging metrics during training. source
- Preprocessing
- Ray Data can distribute heavy preprocessing across CPU nodes so it does not bottleneck GPU training, and Ray Train can split data across workers on the fly. source
- Purpose
- Ray Train distributes model training compute to worker processes across a Ray cluster. source
- Scaling
- The homepage says Ray can scale from a laptop to thousands of GPUs and use heterogeneous GPUs and CPUs with independent scaling. source
- Security
- Ray supports built-in token authentication starting in version 2.52.0, while its security guidance calls for controlled networks and trusted code. source
- Security limitation
- Ray does not provide isolation between jobs or access controls for developers within a cluster; its security guidance recommends separate clusters where workload isolation is required. source
- Support
- The Ray site offers a community Slack, forums, and documentation, and says Anyscale offers hands-on training and expert support. source
- Training workloads
- The homepage describes distributed training for generative AI foundation models, time-series models, and traditional machine-learning models such as XGBoost. source
- Workers and resources
- Ray Train uses a training function, workers, a scaling configuration with CPU or GPU resources, and a Trainer to execute a distributed training job. source
Checked against the maker's pages on 4 Oct 2026.