Cloud-native AI infrastructure for GPU workloads on Kubernetes

TauGrid brings the tau CLI, Kueue queueing, KubeRay orchestration, GPU-node health monitoring, and observability into one stack, spanning data preparation, distributed training, fine-tuning, and inference. Platform teams get an integrated foundation; researchers focus on their code, not Kubernetes plumbing.

Install TauGrid Understand TauGrid Explore examples

One stack, different jobs to be done

Task-oriented paths keep platform bootstrap separate from researcher day-to-day workflows.

Platform engineers stand up and govern the stack

Deploy Kueue, KubeRay, GPU monitoring, and observability, then provide identity, queue, and storage policy.

Platform admin guide →

Researchers run and compare experiments

Connect a repository, submit a smoke test or training run, inspect progress, and preserve results.

Developer guide →

Operators find the first blocked layer

Separate config resolution, Kueue admission, scheduling, GPU health, and model progress.

Troubleshoot a run →

Contributors extend portable capabilities

Choose the owning package, preserve compatibility, and prove the external contract.

Make a compatible change →

From submission to healthy GPUs

A workload submitted with tau run enters a shared Kueue ClusterQueue, receives quota, and is placed by Kubernetes on healthy GPUs. Queue state, placement, and live training evidence stay visible along the way.

An animated TauGrid control room showing jobs from three workspaces sharing a Kueue ClusterQueue while the front job receives GPU quota and Kubernetes places it on two available healthy GPUs.
One shared queue across workspaces Watch jobs enter the ClusterQueue, receive best-effort FIFO admission, and start on healthy GPU capacity.

Across the GPU workload lifecycle

The same stack carries a workload from raw data through to a running endpoint.

An animated TauGrid pipeline carrying a model workload through data preparation, distributed training, fine-tuning, and a ready inference endpoint.
One workload, one continuous record Move from prepared data to a serving endpoint while TauGrid preserves workload evidence.

Clear ownership from cluster to code

TauGrid packages the components needed to admit, run, and observe GPU workloads. Platform teams still operate clusters and capacity; researchers still own workload configuration and model code.

Deploy the GPU workload stack together

Install and upgrade Kueue, KubeRay, GPU health monitoring, and observability as one stack instead of maintaining the integration glue yourself.

Keep workload configuration with the code

Check run settings into the repository, resolve and validate them locally, then submit the rendered workload to Kueue and Kubernetes.

Capture evidence where the run happens

Record experiments, checkpoints, and artifacts locally by default, with optional scalar metric export to ADX.

Built in the open

TauGrid is MIT-licensed and developed in the open on GitHub.

View on GitHub Contribute