Microsoft Open-Sources TauGrid: A Kubernetes-Native Stack for GPU AI Workloads

Platform groups working AI on Kubernetes hardly ever run one factor. They run a queueing system, a distributed runtime, GPU node well being checks, dashboards, and a layer of submission scripts holding all of it collectively. The Azure Kubernetes Service engineering crew open-sourced TauGrid, which collapses that meeting job right into a single Helm set up.

Is it deployable? Sure, TauGrid is MIT licensed, with container photos and Helm charts printed as public OCI artifacts on Microsoft Container Registry. Conditions are a Kubernetes 1.30+ cluster with GPU nodes, kubectl, and Helm 3.0 or later.

What’s TauGrid

TauGrid is a self-hosted platform for working AI workloads on Kubernetes. It combines 5 issues that platform groups often combine by hand: the tau CLI, workload queueing and admission by way of Kueue, Ray cluster orchestration by way of KubeRay, node-level GPU well being monitoring, and cluster and workload observability.

The cut up of accountability is the design level. Platform groups personal workspaces, queues, compute profiles, storage, id, and observability. Researchers work from a repository and the CLI, and submit workloads with out configuring Kubernetes straight. The codebase is written primarily in Go.

How a job strikes by way of it

A workload is described in a tau.yaml file. The GPU coaching instance printed by Microsoft runs a PyTorch job on a single A100:

schema_version: 1
identify: aks-gpu-quickstart
run:
  entrypoint: prepare.py
  workload_kind: rayjob
compute:
  gpus: 1
  staff: 1
  cpus: 16
  reminiscence: 64Gi
runtime:
  picture: mcr.microsoft.com/aks/ai-runtime/ray:py3.12-ray2.56.0-cuda13.0
  pip:
    - torch>=2.4.0

On tau run, TauGrid resolves platform coverage, renders a Kubernetes Job or a KubeRay RayJob, and submits it by way of Kueue. The six phases Microsoft paperwork are submission, queueing, execution, monitoring, restoration, and proof. Restoration covers retry, resume from checkpoint, and failure prognosis. Proof data seize workload metadata, configuration, logs, metrics, checkpoints, and execution historical past, which is what makes a run reproducible and auditable later.

When a number of groups share a cluster, their jobs land in a shared Kueue ClusterQueue. Kueue admits every one on quota and precedence, and Kubernetes locations it on wholesome GPUs.

Interactive explainer

Set up is a Helm chart pulled straight from MCR:

helm set up taugrid 
  oci://mcr.microsoft.com/aks/ai-runtime/helm/taugrid 
  --version 0.4.2 
  --namespace tau-system 
  --create-namespace

First-party photos ship underneath mcr.microsoft.com/aks/ai-runtime/ for Tau, the TauGrid Portal, and the tau core controller. Microsoft advises pinning versioned tags or immutable digests moderately than newest. The CLI installs from GitHub Releases on Linux and macOS, with a PowerShell installer for Home windows amd64; the installer verifies the discharge checksum and doesn’t modify PATH.

Two operational particulars matter for anybody evaluating this outdoors Azure. First, TauGrid sends no telemetry to Microsoft by default, and distant export stays off except an operator configures a vacation spot. Second, some integrations are nonetheless Azure-specific, notably observability by way of Azure Information Explorer. The acknowledged intent is to help cloud and on-premises Kubernetes with out an Azure dependency, and contributions towards which can be open.

Key Takeaways

• Microsoft open-sourced TauGrid on August 28, 2026, underneath the MIT license at Azure/taugrid.
• One Helm set up bundles the tau CLI, Kueue queueing, KubeRay orchestration, GPU well being monitoring, and observability.
• Deployable now on any Kubernetes 1.30+ cluster with GPU nodes, kubectl, and Helm 3.0+.
• Proof data seize config, logs, metrics, and checkpoints, so runs keep reproducible and auditable.
• No telemetry by default, however Azure Information Explorer observability stays Azure-specific for now.

Take a look at the AKS Engineering Blog and Azure/taugrid on GitHub. All credit score goes to the researcher of this mission. Additionally, be at liberty to observe us on Twitter and don’t overlook to affix our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Have to accomplice with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Connect with us