Anyscale launches SIG for Ray on accelerators

News

Anyscale announced a Special Interest Group to expand Ray support for GPUs, TPUs and other accelerators across testing, scheduling and transport.

Anyscale said on Tuesday it is launching a Special Interest Group for Ray on hardware accelerators, a move aimed at making the open-source framework easier to use across a widening mix of GPUs, TPUs and other specialized chips. The group is intended to define a vendor-neutral roadmap for accelerator support as AI systems become more dependent on heterogeneous hardware.

Evidence and context

In the announcement published on Oct. 20, 2026, Anyscale said the SIG will bring together hardware vendors, Ray contributors and application developers to improve accelerator support in Ray across three initial areas: development and testing infrastructure, hardware awareness capabilities, and networking and transport. The company said Ray already supports more than a dozen kinds of accelerators, including NVIDIA and AMD GPUs, TPUs, Trainium and Gaudi.

The proposed work is framed around practical infrastructure concerns rather than product launches. On testing, Anyscale said maintainers need dependable access to real hardware to validate changes, reproduce failures and catch regressions before release. On scheduling and placement, it said Ray needs a more accurate view of accelerator topology, not just counts of devices, so workloads can be assigned to the right hardware. On data movement, the company pointed to Ray Direct Transport, or RDT, which moves tensor data directly between devices using transports such as UCX and RDMA instead of copying through host memory.

Anyscale’s roadmap calls for broader CI coverage, vendor-neutral hardware interfaces, hardened discovery, cloud and platform integration, and validated deployment recipes for high-throughput workloads. The company also said the SIG will define plug-in interfaces, testing and maturity tiers for accelerators in Ray, while vendors and cloud providers contribute device plugins, transports and support for new SKUs.

The announcement follows several recent Anyscale platform updates. On Oct. 2, 2026, the company said GPU pipelining for PyTorch inference entered beta in Anyscale Runtime and requires Ray 2.59 or later. On the same date, Ray 2.59.0 became available as an Anyscale base image. On Oct. 7, 2026, Anyscale on Azure became generally available with native integration on Azure Kubernetes Service.

Ray itself is an open-source framework for scaling Python and AI applications from a laptop to multi-node clusters, with components for distributed execution, resource scheduling and AI workloads such as training, serving and reinforcement learning, according to the project’s documentation. The new SIG is meant to extend that base as accelerator hardware becomes more fragmented and specialized.