SAN FRANCISCO — The PyTorch Foundation, operating under the Linux Foundation, has unveiled Ray, a project framework for artificial intelligence (AI) engineering teams to address the complexity and inefficiency of fragmented, hand-built distributed computing systems, at its annual conference here this week.
Ray, originally developed by Anyscale, provides a unified platform for executing data processing, model training, and serving workloads, scaling seamlessly from single machines to thousands of nodes. Since its origins at UC Berkeley, the project has accumulated more than 39,000 GitHub stars and surpassed 237 million downloads.
“The PyTorch Foundation is committed to fostering an open, interoperable, and production-ready AI ecosystem,” said Matt White, general manager of AI at the Linux Foundation and executive director of the PyTorch Foundation. “By bringing Ray under the PyTorch Foundation umbrella, alongside projects like vLLM and DeepSpeed, we are uniting the critical components needed to build next-generation AI systems.”
Ray tackles three key distributed computing challenges: multimodal data processing across diverse datasets that include text, images, audio, and video; pre-training and post-tuning that scales PyTorch and other machine learning frameworks across thousands of GPUs; and distributed inference that delivers high-throughput, low-latency model serving in production environments.
The framework’s addition creates what foundation leaders describe as an integrated open-source stack. PyTorch handles model development, vLLM manages inference, and Ray executes distributed workloads, allowing development teams to build scalable applications without proprietary infrastructure lock-in.
Chris Aniszczyk, chief technology officer at the Cloud Native Computing Foundation, welcomed the move, noting that Ray and Kubernetes offer complementary capabilities for scaling AI systems through their combined orchestration and distributed compute strengths.
At the same show, Lightning AI unveiled new tools aimed at accelerating distributed training, reinforcement learning, and experimentation for the PyTorch community.
The company’s latest release centers on a purpose-built AI Code Editor integrated into Lightning Studios, featuring domain-specific experts for training, inference, and reinforcement learning tasks. The editor connects directly to Lightning’s GPU Marketplace, enabling developers to provision computing resources without manual cluster configuration.
Additional releases include Lightning Environments for interactive and large-scale training, an official integration with Meta’s Monarch for interactive distributed training, and immediate support for OpenEnv, Meta Platform Inc.’s new open standard for reinforcement learning.
“Our goal is to make every developer in the world a PyTorch developer,” said William Falcon, CEO of Lightning AI and creator of PyTorch Lightning. “Whether you’re training a model on one GPU or hundreds, Lightning gives you the same tight, interactive development loop people love, now supercharged by agents and instantly connected to the compute you need.”
The tools provide access to Lightning’s Model APIs, allowing developers to work with both open and proprietary AI models within a unified workflow, complete with usage tracking, access control, and integrated billing.
