ray-project/ray

Ray is a distributed computing framework that lets you scale Python and machine learning workloads from a single laptop to a large cluster of machines with minimal code changes. The core problem it solves is that modern AI training, inference, and data processing jobs are too compute-intensive for a single machine, but distributing workloads across multiple machines traditionally requires complex infrastructure knowledge that data scientists and ML engineers often lack. Ray addresses this by providing simple Python abstractions. The two most fundamental ones are Tasks, Python functions that run in parallel across available CPU and GPU resources, and Actors, stateful Python objects that persist across multiple calls and can run concurrently. You annotate regular Python functions or classes with Ray decorators, and Ray automatically distributes and parallelizes their execution across whatever hardware is available, whether that's the multiple cores on your laptop or thousands of machines in a cloud cluster. Built on top of this core, Ray ships a set of higher-level libraries for common ML workloads: Ray Data for processing large datasets, Ray Train for distributed model training across frameworks like PyTorch and TensorFlow, Ray Tune for hyperparameter search at scale, RLlib for reinforcement learning, and Ray Serve for deploying models as scalable HTTP endpoints. Someone would use Ray when they have a Python workload that takes too long on one machine, training a large neural network, running a parameter sweep across hundreds of configurations, preprocessing a terabyte-scale dataset, or serving a model that needs to handle thousands of concurrent requests. It works on any cloud provider, on-premise hardware, or Kubernetes. The primary tech stack is Python, with the underlying runtime implemented in C++ for performance. Installation is via pip.