Databricks SDK for JavaScript
    Preparing search index...

    Interface DeploymentSpec

    DeploymentSpec: configuration for one deployment within an AiRuntimeTask. Each entry in AiRuntimeTask.deployments describes a group of nodes that share the same command and compute. Many single-program training algorithms use a single entry where every node runs the same command; role-split workloads (driver + worker, parameter server, separate eval node, etc.) use multiple entries.

    interface DeploymentSpec {
        commandPath?: string;
        compute?: ComputeSpec;
        name?: string;
    }
    Index

    Properties

    commandPath?: string

    Workspace path of the script to run on each node in this deployment. Upload the script to this path and supply the path here. When the task runs, the file at this path is run on each node; if it fails, the task fails with its exit code.

    Example script contents:

    Plain Python:

    python train.py --epochs 10

    Multi-GPU via accelerate:

    accelerate launch train.py --config config.yaml

    Distributed via torchrun:

    torchrun --nproc_per_node=8 train.py

    compute?: ComputeSpec

    Compute resources allocated to each node in this deployment.

    name?: string

    Optional human-readable name for this deployment (for example, driver, worker, param_server). Used for log and UI display. Distinct names are recommended so deployments can be told apart, but uniqueness is not enforced.