Skypilot VM Orchestrator

Orchestrating your pipelines to run on VMs using SkyPilot.

The SkyPilot VM Orchestrator is an integration provided by ZenML that allows you to provision and manage virtual machines (VMs) on any cloud provider supported by the SkyPilot framework. This integration is designed to simplify the process of running machine learning workloads on the cloud, offering cost savings, high GPU availability, and managed execution, We recommend using the SkyPilot VM Orchestrator if you need access to GPUs for your workloads, but don't want to deal with the complexities of managing cloud infrastructure or expensive managed solutions.

When to use it

You should use the SkyPilot VM Orchestrator if:

  • you want to maximize cost savings by leveraging spot VMs and auto-picking the cheapest VM/zone/region/cloud.

  • you want to ensure high GPU availability by provisioning VMs in all zones/regions/clouds you have access to.

  • you don't need a built-in UI of the orchestrator. (You can still use ZenML's Dashboard to view and monitor your pipelines/artifacts.)

  • you're not willing to maintain Kubernetes-based solutions or pay for managed solutions like Sagemaker.

How it works

The orchestrator leverages the SkyPilot framework to handle the provisioning and scaling of VMs. It automatically manages the process of launching VMs for your pipelines, with support for both on-demand and managed spot VMs. While you can select the VM type you want to use, the orchestrator also includes an optimizer that automatically selects the cheapest VM/zone/region/cloud for your workloads. Finally, the orchestrator includes an autostop feature that cleans up idle clusters, preventing unnecessary cloud costs.

You can configure the SkyPilot VM Orchestrator to use a specific VM type, and resources for each step of your pipeline can be configured individually. Read more about how to configure step-specific resources here.

All ZenML pipeline runs are executed using Docker containers within the VMs provisioned by the orchestrator. For that reason, you may need to configure your pipeline settings with docker_run_args=["--gpus=all"] to enable GPU support in the Docker container.

How to deploy it

You don't need to do anything special to deploy the SkyPilot VM Orchestrator. As the SkyPilot integration itself takes care of provisioning VMs, you can simply use the orchestrator as you would any other ZenML orchestrator. However, you will need to ensure that you have the appropriate permissions to provision VMs on your cloud provider of choice and to configure your SkyPilot orchestrator accordingly using the service connectors feature.

The SkyPilot VM Orchestrator currently only supports the AWS, GCP, Azure, Lambda Labs and Kubernetes platforms.

How to use it

To use the SkyPilot VM Orchestrator, you need:

We need first to install the SkyPilot integration for AWS and the AWS connectors extra, using the following commands:

  # Installs dependencies for Skypilot AWS, AWS Container Registry, and S3 Artifact Store
  pip install "zenml[connectors-aws]"
  zenml integration install aws skypilot_aws  # We recommend using the --uv option here

To provision VMs on AWS, your VM Orchestrator stack component needs to be configured to authenticate with AWS Service Connector. To configure the AWS Service Connector, you need to register a new service connector configured with AWS credentials that have at least the minimum permissions required by SkyPilot as documented here.

First, check that the AWS service connector type is available using the following command:

zenml service-connector list-types --type aws
┏━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━┯━━━━━━━━━━━━━━━━━━━━━━━┯━━━━━━━━━━━━━━━━━━┯━━━━━━━┯━━━━━━━━┓
┃         NAME          │ TYPE   │ RESOURCE TYPES        │ AUTH METHODS     │ LOCAL │ REMOTE ┃
┠───────────────────────┼────────┼───────────────────────┼──────────────────┼───────┼────────┨
┃ AWS Service Connector │ 🔶 aws │ 🔶 aws-generic        │ implicit         │ ✅    │ ➖     ┃
┃                       │        │ 📦 s3-bucket          │ secret-key       │       │        ┃
┃                       │        │ 🌀 kubernetes-cluster │ sts-token        │       │        ┃
┃                       │        │ 🐳 docker-registry    │ iam-role         │       │        ┃
┃                       │        │                       │ session-token    │       │        ┃
┃                       │        │                       │ federation-token │       │        ┃
┗━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━┷━━━━━━━━━━━━━━━━━━━━━━━┷━━━━━━━━━━━━━━━━━━┷━━━━━━━┷━━━━━━━━┛

Next, configure a service connector using the CLI or the dashboard with the AWS credentials. For example, the following command uses the local AWS CLI credentials to auto-configure the service connector:

zenml service-connector register aws-skypilot-vm --type aws --region=us-east-1 --auto-configure

This will automatically configure the service connector with the appropriate credentials and permissions to provision VMs on AWS. You can then use the service connector to configure your registered VM Orchestrator stack component using the following command:

# Register the orchestrator
zenml orchestrator register <ORCHESTRATOR_NAME> --flavor vm_aws
# Connect the orchestrator to the service connector
zenml orchestrator connect <ORCHESTRATOR_NAME> --connector aws-skypilot-vm

# Register and activate a stack with the new orchestrator
zenml stack register <STACK_NAME> -o <ORCHESTRATOR_NAME> ... --set

When you are running a pipeline with the Skypilot orchestrator (especially on AWS), we recommend you to use uv as the python package installer:

from zenml import pipeline
from zenml.config.docker_settings import DockerSettings

docker_settings = DockerSettings(python_package_installer="uv")

@pipeline(settings={"docker": docker_settings})
def your_pipeline():
    ...

Additional Configuration

For additional configuration of the Skypilot orchestrator, you can pass Settings depending on which cloud you are using which allows you to configure (among others) the following attributes:

  • instance_type: The instance type to use.

  • cpus: The number of CPUs required for the task. If a string, must be a string of the form '2' or '2+', where the + indicates that the task requires at least 2 CPUs.

  • memory: The amount of memory in GiB required. If a string, must be a string of the form '16' or '16+', where the + indicates that the task requires at least 16 GB of memory.

  • accelerators: The accelerators required. If a string, must be a string of the form 'V100' or 'V100:2', where the :2 indicates that the task requires 2 V100 GPUs. If a dict, must be a dict of the form {'V100': 2} or {'tpu-v2-8': 1}.

  • accelerator_args: Accelerator-specific arguments. For example, {'tpu_vm': True, 'runtime_version': 'tpu-vm-base'} for TPUs.

  • use_spot: Whether to use spot instances. If None, defaults to False.

  • job_recovery: The spot recovery strategy to use for the managed spot to recover the cluster from preemption. Read more about the available strategies here

  • region: The cloud region to use.

  • zone: The cloud zone to use within the region.

  • image_id: The image ID to use. If a string, must be a string of the image id from the cloud, such as AWS: 'ami-1234567890abcdef0', GCP: 'projects/my-project-id/global/images/my-image-name'; Or, a image tag provided by SkyPilot, such as AWS: 'skypilot:gpu-ubuntu-2004'. If a dict, must be a dict mapping from region to image ID.

  • disk_size: The size of the OS disk in GiB.

  • disk_tier: The disk performance tier to use. If None, defaults to 'medium'.

  • cluster_name: Name of the cluster to create/reuse. If None, auto-generate a name. SkyPilot uses term cluster to refer to a group or a single VM that are provisioned to execute the task. The cluster name is used to identify the cluster and to determine whether to reuse an existing cluster or create a new one.

  • retry_until_up: Whether to retry launching the cluster until it is up.

  • idle_minutes_to_autostop: Automatically stop the cluster after this many minutes of idleness, i.e., no running or pending jobs in the cluster's job queue. Idleness gets reset whenever setting-up/running/pending jobs are found in the job queue. Setting this flag is equivalent to running sky.launch(..., detach_run=True, ...) and then sky.autostop(idle_minutes=<minutes>). If not set, the cluster will not be autostopped.

  • down: Tear down the cluster after all jobs finish (successfully or abnormally). If idle_minutes_to_autostop is also set, the cluster will be torn down after the specified idle time. Note that if errors occur during provisioning/data syncing/setting up, the cluster will not be torn down for debugging purposes.

  • stream_logs: If True, show the logs in the terminal as they are generated while the cluster is running.

  • docker_run_args: Additional arguments to pass to the docker run command. For example, ['--gpus=all'] to use all GPUs available on the VM.

  • ports: Ports to expose. Could be an integer, a range, or a list of integers and ranges. All ports will be exposed to the public internet.

  • labels: Labels to apply to instances as key-value pairs. These are mapped to cloud-specific implementations (instance tags in AWS, instance labels in GCP, etc.).

  • any_of: List of candidate resources to try in order of preference based on cost (determined by the SkyPilot optimizer).

  • ordered: List of candidate resources to try in the specified order.

  • workdir: Working directory on the local machine to sync to the VM. This is synced to ~/sky_workdir inside the VM.

  • task_name: Human-readable task name shown in SkyPilot for display purposes.

  • file_mounts: File and storage mounts configuration to make local or cloud storage paths available inside the remote cluster.

  • envs: Environment variables for the task. Accessible in the VMs that Skypilot launches, not in Docker continaers that the steps and pipeline is running on.

  • task_settings: Dictionary of arbitrary settings forwarded to sky.Task(). This allows passing future parameters added by SkyPilot without requiring updates to ZenML.

  • resources_settings: Dictionary of arbitrary settings forwarded to sky.Resources(). This allows passing future parameters added by SkyPilot without requiring updates to ZenML.

  • launch_settings: Dictionary of arbitrary settings forwarded to sky.launch(). This allows passing future parameters added by SkyPilot without requiring updates to ZenML.

The following code snippets show how to configure the orchestrator settings for each cloud provider:

Code Example:

from zenml.integrations.skypilot_aws.flavors.skypilot_orchestrator_aws_vm_flavor import SkypilotAWSOrchestratorSettings

skypilot_settings = SkypilotAWSOrchestratorSettings(
    cpus="2",
    memory="16",
    accelerators="V100:2",
    accelerator_args={"tpu_vm": True, "runtime_version": "tpu-vm-base"},
    use_spot=True,
    job_recovery={
        "strategy": "failover",
        "max_restarts_on_errors": 3,
    },
    region="us-west-1",
    zone="us-west1-a",
    image_id="ami-1234567890abcdef0",
    disk_size=100,
    disk_tier="high",
    cluster_name="my_cluster",
    retry_until_up=True,
    idle_minutes_to_autostop=60,
    down=True,
    stream_logs=True,
    docker_run_args=["--gpus=all"]
)


@pipeline(
    settings={
        "orchestrator": skypilot_settings
    }
)

One of the key features of the SkyPilot VM Orchestrator is the ability to run each step of a pipeline on a separate VM with its own specific settings. This allows for fine-grained control over the resources allocated to each step, ensuring that each part of your pipeline has the necessary compute power while optimizing for cost and efficiency.

Configuring Step-Specific Resources

The SkyPilot VM Orchestrator allows you to configure resources for each step individually. This means you can specify different VM types, CPU and memory requirements, and even use spot instances for certain steps while using on-demand instances for others.

If no step-specific settings are specified, the orchestrator will use the resources specified in the orchestrator settings for each step and run the entire pipeline in one VM. If step-specific settings are specified, an orchestrator VM will be spun up first, which will subsequently spin out new VMs dependent on the step settings. You can disable this behavior by setting the disable_step_based_settings parameter to True in the orchestrator configuration, using the following command:

zenml orchestrator update <ORCHESTRATOR_NAME> --disable_step_based_settings=True

Here's an example of how to configure specific resources for a step for the AWS cloud:

from zenml.integrations.skypilot_aws.flavors.skypilot_orchestrator_aws_vm_flavor import SkypilotAWSOrchestratorSettings

# Settings for a specific step that requires more resources
high_resource_settings = SkypilotAWSOrchestratorSettings(
    instance_type='t2.2xlarge',
    cpus=8,
    memory=32,
    use_spot=False,
    region='us-east-1',
    # ... other settings
)

@step(settings={"orchestrator": high_resource_settings})
def my_resource_intensive_step():
    # Step implementation
    pass

By using the settings parameter, you can tailor the resources for each step according to its specific needs. This flexibility allows you to optimize your pipeline execution for both performance and cost.

Check out the SDK docs for a full list of available attributes and this docs page for more information on how to specify settings.

ZenML Scarf

Last updated

Was this helpful?