Ray

Ray is an open source framework for distributed computing, positioned as an AI compute engine for scaling machine learning and AI workloads. It is native to Python and is developed by Anyscale. It lets developers take Python code from a laptop to thousands of GPUs, using CPUs and GPUs of different kinds in the same pipeline, with independent scaling of each part of a workload. The site lists the workloads Ray supports: parallel Python code for uses such as simulation and backtesting, multimodal data processing of images, video and audio, distributed model training including generative AI foundation models and classic models such as XGBoost, model serving through Ray Serve with fractional resources, batch inference, reinforcement learning through RLlib, generative AI workflows including retrieval augmented generation, and LLM inference both online and batch. Ray works with the developer's choice of training framework. The audience is machine learning and platform engineers who need to run training, data processing and inference at scale without building their own orchestration. The framework itself is free to use. The site promotes a $100 credit to try Ray on Anyscale's managed platform, and no pricing page is published on ray.io.