Skip to content

0.56.2
Compare
Choose a tag to compare
@safoinme safoinme released this 25 Mar 22:27
· 545 commits to main since this release
0.56.2
68bcb3b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Learn about vigilant mode.

This release introduces a wide array of new features, enhancements, and bug fixes, with a strong emphasis on elevating the user experience and streamlining machine
learning workflows. Most notably, you can now deploy models using Hugging Face inference endpoints thanks to an open-source community contribution of this model deployer stack component!

Note that 0.56.0 and 0.56.1 were yanked and removed from PyPI due to an issue with the
alembic versions + migration which could affect the database state. This release
fixes that issue.

This release also comes with a breaking change to the services
architecture.

Breaking Change

A significant change in this release is the migration of the Service (ZenML's technical term for deployment)
registration and deployment from local or remote environments to the ZenML server.
This change will be reflected in an upcoming tab in the dashboard which will
allow users to explore and see the deployed models in the dashboard with their live
status and metadata. This architectural shift also simplifies the model deployer
abstraction and streamlines the model deployment process for users by moving from
limited built-in steps to a more documented and flexible approach.

Important note: If you have models that you previously deployed with ZenML, you might
want to redeploy them to have them stored in the ZenML server and tracked by ZenML,
ensuring they appear in the dashboard.

Additionally, the find_model_server method now retrieves models (services) from the
ZenML server instead of local or remote deployment environments. As a result, any
usage of find_model_server will only return newly deployed models stored in the server.

It is also no longer recommended to call service functions like service.start().
Instead, use model_deployer.start_model_server(service_id), which will allow ZenML
to update the changed status of the service in the server.

Starting a service

Old syntax:

from zenml import pipeline, 
from zenml.integrations.bentoml.services.bentoml_deployment import BentoMLDeploymentService

@step
def predictor(
    service: BentoMLDeploymentService,
) -> None:
    # starting the service
    service.start(timeout=10)

New syntax:

from zenml import pipeline
from zenml.integrations.bentoml.model_deployers import BentoMLModelDeployer
from zenml.integrations.bentoml.services.bentoml_deployment import BentoMLDeploymentService

@step
def predictor(
    service: BentoMLDeploymentService,
) -> None:
    # starting the service
    model_deployer = BentoMLModelDeployer.get_active_model_deployer()
    model_deployer.start_model_server(service_id=service.service_id, timeout=10)

Enabling continuous deployment

Instead of replacing the parameter that was used in the deploy_model method to replace the
existing service (if it matches the exact same pipeline name and step name without
taking into accounts other parameters or configurations), we now have a new parameter,
continuous_deployment_mode, that allows you to enable continuous deployment for
the service. This will ensure that the service is updated with the latest version
if it's on the same pipeline and step and the service is not already running. Otherwise,
any new deployment with different configurations will create a new service.

from zenml import pipeline, step, get_step_context
from zenml.client import Client

@step
def deploy_model() -> Optional[MLFlowDeploymentService]:
    # Deploy a model using the MLflow Model Deployer
    zenml_client = Client()
    model_deployer = zenml_client.active_stack.model_deployer
    mlflow_deployment_config = MLFlowDeploymentConfig(
        name: str = "mlflow-model-deployment-example",
        description: str = "An example of deploying a model using the MLflow Model Deployer",
        pipeline_name: str = get_step_context().pipeline_name,
        pipeline_step_name: str = get_step_context().step_name,
        model_uri: str = "runs:/<run_id>/model" or "models:/<model_name>/<model_version>",
        model_name: str = "model",
        workers: int = 1
        mlserver: bool = False
        timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT
    )
    service = model_deployer.deploy_model(mlflow_deployment_config, continuous_deployment_mode=True)
    logger.info(f"The deployed service info: {model_deployer.get_model_server_info(service)}")
    return service

Major Features and Enhancements:

  • A new Huggingface Model Deployer has been introduced, allowing you to seamlessly
    deploy your Huggingface models using ZenML. (Thank you so much @dudeperf3ct for the contribution!)
  • Faster Integration and Dependency Management ZenML now leverages the uv library,
    significantly improving the speed of integration installations and dependency management,
    resulting in a more streamlined and efficient workflow.
  • Enhanced Logging and Status Tracking Logging have been improved, providing better
    visibility into the state of your ZenML services.
  • Improved Artifact Store Isolation: ZenML now prevents unsafe operations that access
    data outside the scope of the artifact store, ensuring better isolation and security.
  • Adding admin user notion for the user accounts and added protection to certain operations
    performed via the REST interface to ADMIN-allowed only.
  • Rate limiting for login API to prevent abuse and protect the server from potential
    security threats.
  • The LLM template is now supported in ZenML, allowing you to use the LLM template
    for your pipelines.

🥳 Community Contributions 🥳

We'd like to give a special thanks to @dudeperf3ct he contributed to this release
by introducing the Huggingface Model Deployer. We'd also like to thank @moesio-f
for their contribution to this release by adding a new attribute to the Kaniko image builder.
Additionally, we'd like to thank @christianversloot for his contributions to this release.

What's Changed

New Contributors

Full Changelog: 0.55.5...0.56.2

Contributors

strickvl, stefannica, and 10 other contributors
Assets 2
Loading