# Anyscale Forms Ray Group to Improve Support Across AI Chips By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/36354 Published: 2026-10-21T04:58:18.000Z Updated: 2026-10-21T04:58:18.000Z Section: Technology > The group plans to work on hardware testing, job scheduling and data transfers across different accelerators. Anyscale launched a special interest group for Ray on Oct. 20 to improve testing, job placement and data transfers as AI developers use a wider mix of specialized chips. Ray is an open-source framework that distributes Python and AI applications across clusters of computers. It already supports more than a dozen types of accelerators, including NVIDIA and AMD GPUs, TPUs, Trainium and Gaudi. The group aims to bring hardware vendors, Ray contributors and application developers together to improve how Ray works with those devices. Its initial work covers three areas: development and testing, awareness of hardware configurations, and movement of data between devices. The plans address practical problems that arise when software must run across chips with different capabilities and connections. The group will also work on common plug-in interfaces so support for additional hardware can be developed within Ray. For testing, the group wants contributors to have access to physical chips. That would let them reproduce failures, check changes on the affected hardware and catch problems before a release. The roadmap also calls for broader automated checks across devices and workloads, giving maintainers more ways to assess an integration as it develops. For scheduling, Ray needs a clearer view of how accelerators are connected and grouped. A device count alone does not describe where a job’s work should run. The planned changes would help Ray place work on suitable hardware and improve how it discovers the devices available in a cluster. The third area is data movement. Ray Direct Transport can send tensor data directly between devices without copying it through host memory. The group plans to extend that capability as it develops transport methods for a broader range of accelerators. Also on Oct. 20, [Anyscale and Google Cloud introduced native PyTorch support for Ray on Google TPUs](). That integration assigns each Ray worker to one TPU chip and allows a job’s workers to be scheduled together on the same TPU slice. The new group’s proposed work spans multiple accelerator types rather than that specific TPU setup. The group plans to define plug-in interfaces, continuous integration testing and maturity tiers for accelerator support. Hardware vendors and cloud providers are expected to contribute and maintain device plug-ins and transports as that work moves forward. ## Links in this article - [Anyscale and Google Cloud introduced native PyTorch support for Ray on Google TPUs](https://www.tokenpost.com/news/technology/36258)