Getting a training job onto a GPU cluster usually means learning a platform team’s home-made conventions first. Microsoft’s answer, published as open source, is to put those conventions into one product.
TauGrid targets Kubernetes clusters with GPU nodes. It bundles the pieces teams otherwise assemble by hand: Kueue for queuing and quota, KubeRay for orchestration, health monitoring for GPU nodes and observability, installed from a single Helm chart with ownership boundaries drawn around each part.
A researcher describes a job in a yaml file and submits it with the tau command. The tool validates the file and creates either a Kubernetes Job or a RayJob, after which Kueue places it by remaining quota and priority. Status, logs and checkpoints are tracked while the job runs, and evidence is stored so a result can be reproduced or a failure explained. A crashed job resumes from its last checkpoint.
Platform engineers keep what they need: workspaces, compute profiles, storage, identity and observability sit behind a supported interface rather than a wiki page.
The roadmap is candid about gaps. Multi-tenant workspaces, role-based access and quotas, PyTorch DDP and FSDP, DeepSpeed and LoRA workflows, dataset lifecycle management and multi-cluster execution are all planned. The code is mostly Go and developed in the open inside the Azure ecosystem. Kubernetes 1.30 or newer with GPU nodes, kubectl and Helm 3 are the requirements.
