Getting started

TauGrid is an open, self-hosted AI compute platform for teams that run AI workloads on Kubernetes. It gives developers a consistent path from a repository to scheduled compute and saved results. It gives platform teams a consistent way to provide workspaces, share GPUs, apply policy, and understand what is happening on the cluster.

The idea behind TauGrid

Developers should be able to submit training, tuning, batch inference, and serving workloads without rebuilding the same cluster integration for every project. Platform teams should be able to operate shared compute without creating a separate workflow for every research team.

TauGrid connects those needs:

repository -> tau run -> governed workspace -> shared compute -> saved evidence

The repository describes the workload. The workspace supplies the namespace, queue, identity, storage, and platform policy. Kueue admits work when quota is available. Kubernetes places pods on healthy capacity. TauGrid keeps the developer experience consistent across that path.

Design principles

• Build on Kubernetes. Cluster state, scheduling, identity, and workload resources remain visible through standard Kubernetes APIs.
• Keep workload intent with the code. A checked-in target describes how a project runs, which makes submissions repeatable across developers and environments.
• Share accelerators through queues. Kueue admission and platform-owned quotas make waiting, priority, and GPU allocation explicit.
• Give each system a clear job. TauGrid handles the workflow, Kueue handles admission, Kubernetes handles placement, and Ray or the application handles execution.
• Save evidence as part of the run. Status, logs, metrics, and artifacts help teams compare work, diagnose failures, and promote successful results.
• Support the infrastructure teams already operate. TauGrid uses Kubernetes contracts for compute, storage, networking, and identity across cloud and datacenter environments.

Choose your path

• Developers and researchers: start with the Developer guide after receiving a workspace.
• Platform owners: use the Platform admin guide to install TauGrid and prepare that workspace.

Continue with What is TauGrid? for the product model, Architecture for component ownership, or Core technologies for Ray, Kueue, adx-mon, and Stellar.