DGXCloudExecutor

Launch distributed jobs on NVIDIA DGX Cloud via the Run:ai API.

Prerequisites

`DGXCloudExecutor` is currently only supported when launching experiments *from a pod running on the DGX Cloud cluster itself*. The launching pod must have access to a Persistent Volume Claim (PVC), and the same PVC must be mounted by the launched job.

You need:

• Access to a DGX Cloud cluster and a Run:ai project
• A Run:ai application ID and secret (create one in the Run:ai console under Application credentials)
• A PVC accessible from both the launching pod and the job pods
• NEMORUN_HOME pointing to a path on the PVC

Executor configuration

import nemo_run as run

executor = run.DGXCloudExecutor(
    base_url="https://<cluster-name>.<domain>/api/v1",  # Run:ai API endpoint
    app_id="YOUR_RUNAI_APP_ID",
    app_secret="YOUR_RUNAI_APP_SECRET",
    project_name="YOUR_RUNAI_PROJECT_NAME",
    container_image="nvcr.io/nvidia/pytorch:24.05-py3",
    nodes=1,
    gpus_per_node=8,
    pvcs=[{
        "name": "your-pvc-k8s-name",    # Kubernetes PVC name
        "path": "/your_custom_path",     # mount path inside the container
    }],
    env_vars={"PYTHONUNBUFFERED": "1"},
)

Key parameters:

Parameter	Description
base_url	Run:ai REST API base URL
app_id / app_secret	Run:ai application credentials
project_name	Run:ai project to submit jobs to
container_image	Container image URI
nodes	Number of nodes
gpus_per_node	GPUs per node
pvcs	List of PVC mounts (name + path)

E2E workflow

import nemo_run as run

task = run.Script("python train.py --lr=3e-4 --max-steps=500")

executor = run.DGXCloudExecutor(
    base_url="https://my-cluster.example.com/api/v1",
    app_id="my-app-id",
    app_secret="my-app-secret",
    project_name="my-project",
    container_image="nvcr.io/nvidia/pytorch:24.05-py3",
    nodes=1,
    gpus_per_node=8,
    pvcs=[{"name": "my-pvc", "path": "/workspace"}],
)

with run.Experiment("my-experiment") as exp:
    exp.add(task, executor=executor, name="training")
    exp.run(detach=True)

# Later — reconnect and check status
experiment = run.Experiment.from_id("my-experiment_<id>")
experiment.status()
experiment.logs("training")

Advanced options

Package code from git

executor = run.DGXCloudExecutor(
    ...,
    packager=run.GitArchivePackager(subpath="src"),
)

For a complete end-to-end NeMo example see the NVIDIA DGX Cloud NeMo E2E Workflow.