Use the Databricks CLI with AI Runtime
The Databricks CLI commands for AI Runtime are in Public Preview. A workspace admin must enable the AI Runtime preview before you can submit workloads. See Requirements.
The databricks air command group submits and manages distributed training workloads on AI Runtime, the on-demand serverless GPU compute platform. It uses YAML-based job configuration, integrates with MLflow, and supports workspace-based and git-based code workflows.
The Databricks CLI commands for AI Runtime meet the same compliance standards as AI Runtime workloads. For the supported standards and exceptions, see compliance security profile.
Migrate from the Python CLI
The Databricks CLI replaces the Python-based air CLI distributed as the databricks-air package. The Python CLI does not receive new features. Install the latest Databricks CLI to use the latest AI Runtime features. The Databricks CLI supports macOS, Linux, and Windows.
Some command names and flags differ, so migration requires more than adding the databricks prefix. See command differences.
When to use the CLI
Use the Databricks CLI with AI Runtime when you want to:
- Submit GPU training workloads from your laptop and code editor without opening a notebook.
- Define training jobs declaratively in YAML so they can be checked into source control.
- Run long-running training jobs or multi-node distributed training — workloads that need to outlive an interactive session or span more than one node.
For the in-notebook Python API (@distributed and @ray_launch), see Distributed training in notebooks instead.
To work interactively on a single GPU node over SSH — from your terminal or IDE, rather than submitting a workload — use databricks ssh connect. See Connect to Databricks using an SSH tunnel.
Get started
Install the latest Databricks CLI and configure authentication. See Install or update the Databricks CLI and Authentication for the Databricks CLI.
Then follow the quickstart to submit your first workload. For commands and flags, see air command group.