Raydial is an independent Ray distribution and managed runtime for neoclouds and enterprises. Run production Ray without handing your platform to a competitor.
pip install raydial
Start with the Ray Workflows drop-in. One import change.
With Anyscale now part of Nscale, the default managed path for Ray belongs to a neocloud. If you run GPUs for a living, that is a supplier you also compete with.
Ship managed Ray to your customers under your brand, on your hardware, without routing them through a rival's control plane.
Self-hosted or cloud-hosted, Raydial runs inside your boundary with upgrades, patches and support handled for you.
Raydial tracks open-source Ray. No proprietary fork to get locked into, and the improvements go back upstream.
Ray Workflows was deprecated upstream in mid-2025 (PR #53612). Teams that built durable pipelines on it were left without a path forward.
Raydial ships a drop-in replacement library. Your workflow code stays the same; only the import changes.
- from ray import workflow + from raydial import workflow # everything below is unchanged @ray.remote def fetch(url): ... @ray.remote def train(data): ... dag = train.bind(fetch.bind("s3://corpus")) workflow.run(dag, workflow_id="nightly-train")
A downstream Ray distribution with the operational layer neoclouds and platform teams need: provisioning, upgrades, tenancy and observability, packaged so you can run it as a service.
Raydial is also the runtime that JAGNAUT invokes at apply-time, so declared AI infrastructure lands on a managed Ray cluster.
A distribution is only as good as its relationship with upstream. Raydial's core changes are proposed to the Ray project first.
A Ray Enhancement Proposal to open a seam at Plasma, so the object store can be swapped for storage tuned to the hardware it runs on.
The upstream change that removed Ray Workflows. The Raydial drop-in exists so the teams depending on it are not stranded.