Workbench
Workbench is where you build, train, and serve machine-learning models on Vantage. It organizes everything into five sections, Workspace, Develop, Compute, Train, and Serve, all sharing the same compute, storage, and cost envelope.
What you'll find inside
- Sessions: Interactive notebook environments, Jupyter, VS Code, RStudio, pinned to GPU pools and your team's storage.
- Presets: Reusable templates that define which IDE, images, compute sizes, and storage a session gets.
- Cloud Shell: Browser-based terminal sessions with optional Slurm access.
- Remote Desktop: VNC-based remote desktop sessions for full GUI access to GPU nodes.
- PVC Viewer: Browser-based file browser for PersistentVolumeClaims.
- Compute Profiles: The reusable shape of your compute: GPU type, count, autoscaling bounds, instance class.
- Training Jobs: Distributed training on PyTorch, DeepSpeed, or MLX runtimes. Retry, suspend, resume.
- Pipelines: Multi-step DAGs that orchestrate ingestion, training, evaluation, and deployment.
- Sweeps: Hyperparameter search with Bayesian, grid, or random algorithms, tracked end-to-end.
- Models: Versioned model catalog. Pull from HuggingFace or your own training runs, then deploy to an endpoint.
- Endpoints: Inference services with autoscaling, canary rollouts, and authenticated URLs.
- Observability: Cluster-wide rollups of utilization, spend, idle GPU hours, and live alerts.
- Keyboard shortcuts: Quick reference for Workbench keyboard shortcuts.