Vllm

zenml.integrations.vllm

Initialization for the ZenML vLLM integration.

Attributes

VLLM = 'vllm' module-attribute

VLLM_MODEL_DEPLOYER = 'vllm' module-attribute

logger = get_logger(__name__) module-attribute

Classes

Flavor

Class for ZenML Flavors.

Attributes

config_class: Type[StackComponentConfig] abstractmethod property

Returns StackComponentConfig config class.

Returns:

Type	Description
Type[StackComponentConfig]	The config class.

config_schema: Dict[str, Any] property

The config schema for a flavor.

Returns:

Type	Description
Dict[str, Any]	The config schema.

docs_url: Optional[str] property

A url to point at docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor docs url.

implementation_class: Type[StackComponent] abstractmethod property

Implementation class for this flavor.

Returns:

Type	Description
Type[StackComponent]	The implementation class for this flavor.

logo_url: Optional[str] property

A url to represent the flavor in the dashboard.

Returns:

Type	Description
Optional[str]	The flavor logo.

name: str abstractmethod property

The flavor name.

Returns:

Type	Description
str	The flavor name.

sdk_docs_url: Optional[str] property

A url to point at SDK docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor SDK docs url.

service_connector_requirements: Optional[ServiceConnectorRequirements] property

Service connector resource requirements for service connectors.

Specifies resource requirements that are used to filter the available service connector types that are compatible with this flavor.

Returns:

Type	Description
Optional[ServiceConnectorRequirements]	Requirements for compatible service connectors, if a service
Optional[ServiceConnectorRequirements]	connector is required for this flavor.

type: StackComponentType abstractmethod property

The stack component type.

Returns:

Type	Description
StackComponentType	The stack component type.

Functions

from_model(flavor_model: FlavorResponse) -> Flavor classmethod

Loads a flavor from a model.

Parameters:

Name	Type	Description	Default
flavor_model	FlavorResponse	The model to load from.	required

Raises:

Type	Description
CustomFlavorImportError	If the custom flavor can't be imported.
ImportError	If the flavor can't be imported.

Returns:

Type	Description
Flavor	The loaded flavor.

Source code in src/zenml/stack/flavor.py

@classmethod
def from_model(cls, flavor_model: FlavorResponse) -> "Flavor":
    """Loads a flavor from a model.

    Args:
        flavor_model: The model to load from.

    Raises:
        CustomFlavorImportError: If the custom flavor can't be imported.
        ImportError: If the flavor can't be imported.

    Returns:
        The loaded flavor.
    """
    try:
        flavor = source_utils.load(flavor_model.source)()
    except (ModuleNotFoundError, ImportError, NotImplementedError) as err:
        if flavor_model.is_custom:
            flavor_module, _ = flavor_model.source.rsplit(".", maxsplit=1)
            expected_file_path = os.path.join(
                source_utils.get_source_root(),
                flavor_module.replace(".", os.path.sep),
            )
            raise CustomFlavorImportError(
                f"Couldn't import custom flavor {flavor_model.name}: "
                f"{err}. Make sure the custom flavor class "
                f"`{flavor_model.source}` is importable. If it is part of "
                "a library, make sure it is installed. If "
                "it is a local code file, make sure it exists at "
                f"`{expected_file_path}.py`."
            )
        else:
            raise ImportError(
                f"Couldn't import flavor {flavor_model.name}: {err}"
            )
    return cast(Flavor, flavor)

generate_default_docs_url() -> str

Generate the doc urls for all inbuilt and integration flavors.

Note that this method is not going to be useful for custom flavors, which do not have any docs in the main zenml docs.

Returns:

Type	Description
str	The complete url to the zenml documentation

Source code in src/zenml/stack/flavor.py

def generate_default_docs_url(self) -> str:
    """Generate the doc urls for all inbuilt and integration flavors.

    Note that this method is not going to be useful for custom flavors,
    which do not have any docs in the main zenml docs.

    Returns:
        The complete url to the zenml documentation
    """
    from zenml import __version__

    component_type = self.type.plural.replace("_", "-")
    name = self.name.replace("_", "-")

    try:
        is_latest = is_latest_zenml_version()
    except RuntimeError:
        # We assume in error cases that we are on the latest version
        is_latest = True

    if is_latest:
        base = "https://docs.zenml.io"
    else:
        base = f"https://zenml-io.gitbook.io/zenml-legacy-documentation/v/{__version__}"
    return f"{base}/stack-components/{component_type}/{name}"

generate_default_sdk_docs_url() -> str

Generate SDK docs url for a flavor.

Returns:

Type	Description
str	The complete url to the zenml SDK docs

Source code in src/zenml/stack/flavor.py

def generate_default_sdk_docs_url(self) -> str:
    """Generate SDK docs url for a flavor.

    Returns:
        The complete url to the zenml SDK docs
    """
    from zenml import __version__

    base = f"https://sdkdocs.zenml.io/{__version__}"

    component_type = self.type.plural

    if "zenml.integrations" in self.__module__:
        # Get integration name out of module path which will look something
        #  like this "zenml.integrations.<integration>....
        integration = self.__module__.split(
            "zenml.integrations.", maxsplit=1
        )[1].split(".")[0]

        return (
            f"{base}/integration_code_docs"
            f"/integrations-{integration}/#{self.__module__}"
        )

    else:
        return (
            f"{base}/core_code_docs/core-{component_type}/"
            f"#{self.__module__}"
        )

to_model(integration: Optional[str] = None, is_custom: bool = True) -> FlavorRequest

Converts a flavor to a model.

Parameters:

Name	Type	Description	Default
integration	Optional[str]	The integration to use for the model.	None
is_custom	bool	Whether the flavor is a custom flavor.	True

Returns:

Type	Description
FlavorRequest	The model.

Source code in src/zenml/stack/flavor.py

def to_model(
    self,
    integration: Optional[str] = None,
    is_custom: bool = True,
) -> FlavorRequest:
    """Converts a flavor to a model.

    Args:
        integration: The integration to use for the model.
        is_custom: Whether the flavor is a custom flavor.

    Returns:
        The model.
    """
    connector_requirements = self.service_connector_requirements
    connector_type = (
        connector_requirements.connector_type
        if connector_requirements
        else None
    )
    resource_type = (
        connector_requirements.resource_type
        if connector_requirements
        else None
    )
    resource_id_attr = (
        connector_requirements.resource_id_attr
        if connector_requirements
        else None
    )

    model = FlavorRequest(
        name=self.name,
        type=self.type,
        source=source_utils.resolve(self.__class__).import_path,
        config_schema=self.config_schema,
        connector_type=connector_type,
        connector_resource_type=resource_type,
        connector_resource_id_attr=resource_id_attr,
        integration=integration,
        logo_url=self.logo_url,
        docs_url=self.docs_url,
        sdk_docs_url=self.sdk_docs_url,
        is_custom=is_custom,
    )
    return model

Integration

Base class for integration in ZenML.

Functions

activate() -> None classmethod

Abstract method to activate the integration.

Source code in src/zenml/integrations/integration.py

@classmethod
def activate(cls) -> None:
    """Abstract method to activate the integration."""

check_installation() -> bool classmethod

Method to check whether the required packages are installed.

Returns:

Type	Description
bool	True if all required packages are installed, False otherwise.

Source code in src/zenml/integrations/integration.py

@classmethod
def check_installation(cls) -> bool:
    """Method to check whether the required packages are installed.

    Returns:
        True if all required packages are installed, False otherwise.
    """
    for r in cls.get_requirements():
        try:
            # First check if the base package is installed
            dist = pkg_resources.get_distribution(r)

            # Next, check if the dependencies (including extras) are
            # installed
            deps: List[Requirement] = []

            _, extras = parse_requirement(r)
            if extras:
                extra_list = extras[1:-1].split(",")
                for extra in extra_list:
                    try:
                        requirements = dist.requires(extras=[extra])  # type: ignore[arg-type]
                    except pkg_resources.UnknownExtra as e:
                        logger.debug(f"Unknown extra: {str(e)}")
                        return False
                    deps.extend(requirements)
            else:
                deps = dist.requires()

            for ri in deps:
                try:
                    # Remove the "extra == ..." part from the requirement string
                    cleaned_req = re.sub(
                        r"; extra == \"\w+\"", "", str(ri)
                    )
                    pkg_resources.get_distribution(cleaned_req)
                except pkg_resources.DistributionNotFound as e:
                    logger.debug(
                        f"Unable to find required dependency "
                        f"'{e.req}' for requirement '{r}' "
                        f"necessary for integration '{cls.NAME}'."
                    )
                    return False
                except pkg_resources.VersionConflict as e:
                    logger.debug(
                        f"Package version '{e.dist}' does not match "
                        f"version '{e.req}' required by '{r}' "
                        f"necessary for integration '{cls.NAME}'."
                    )
                    return False

        except pkg_resources.DistributionNotFound as e:
            logger.debug(
                f"Unable to find required package '{e.req}' for "
                f"integration {cls.NAME}."
            )
            return False
        except pkg_resources.VersionConflict as e:
            logger.debug(
                f"Package version '{e.dist}' does not match version "
                f"'{e.req}' necessary for integration {cls.NAME}."
            )
            return False

    logger.debug(
        f"Integration {cls.NAME} is installed correctly with "
        f"requirements {cls.get_requirements()}."
    )
    return True

flavors() -> List[Type[Flavor]] classmethod

Abstract method to declare new stack component flavors.

Returns:

Type	Description
List[Type[Flavor]]	A list of new stack component flavors.

Source code in src/zenml/integrations/integration.py

@classmethod
def flavors(cls) -> List[Type[Flavor]]:
    """Abstract method to declare new stack component flavors.

    Returns:
        A list of new stack component flavors.
    """
    return []

get_requirements(target_os: Optional[str] = None, python_version: Optional[str] = None) -> List[str] classmethod

Method to get the requirements for the integration.

Parameters:

Name	Type	Description	Default
target_os	Optional[str]	The target operating system to get the requirements for.	None
python_version	Optional[str]	The Python version to use for the requirements.	None

Returns:

Type	Description
List[str]	A list of requirements.

Source code in src/zenml/integrations/integration.py

@classmethod
def get_requirements(
    cls,
    target_os: Optional[str] = None,
    python_version: Optional[str] = None,
) -> List[str]:
    """Method to get the requirements for the integration.

    Args:
        target_os: The target operating system to get the requirements for.
        python_version: The Python version to use for the requirements.

    Returns:
        A list of requirements.
    """
    return cls.REQUIREMENTS

get_uninstall_requirements(target_os: Optional[str] = None) -> List[str] classmethod

Method to get the uninstall requirements for the integration.

Parameters:

Name	Type	Description	Default
target_os	Optional[str]	The target operating system to get the requirements for.	None

Returns:

Type	Description
List[str]	A list of requirements.

Source code in src/zenml/integrations/integration.py

@classmethod
def get_uninstall_requirements(
    cls, target_os: Optional[str] = None
) -> List[str]:
    """Method to get the uninstall requirements for the integration.

    Args:
        target_os: The target operating system to get the requirements for.

    Returns:
        A list of requirements.
    """
    ret = []
    for each in cls.get_requirements(target_os=target_os):
        is_ignored = False
        for ignored in cls.REQUIREMENTS_IGNORED_ON_UNINSTALL:
            if each.startswith(ignored):
                is_ignored = True
                break
        if not is_ignored:
            ret.append(each)
    return ret

plugin_flavors() -> List[Type[BasePluginFlavor]] classmethod

Abstract method to declare new plugin flavors.

Returns:

Type	Description
List[Type[BasePluginFlavor]]	A list of new plugin flavors.

Source code in src/zenml/integrations/integration.py

@classmethod
def plugin_flavors(cls) -> List[Type["BasePluginFlavor"]]:
    """Abstract method to declare new plugin flavors.

    Returns:
        A list of new plugin flavors.
    """
    return []

VLLMIntegration

Bases: Integration

Definition of vLLM integration for ZenML.

Functions

activate() -> None classmethod

Activates the integration.

Source code in src/zenml/integrations/vllm/__init__.py

@classmethod
def activate(cls) -> None:
    """Activates the integration."""
    from zenml.integrations.vllm import services

flavors() -> List[Type[Flavor]] classmethod

Declare the stack component flavors for the vLLM integration.

Returns:

Type	Description
List[Type[Flavor]]	List of stack component flavors for this integration.

Source code in src/zenml/integrations/vllm/__init__.py

@classmethod
def flavors(cls) -> List[Type[Flavor]]:
    """Declare the stack component flavors for the vLLM integration.

    Returns:
        List of stack component flavors for this integration.
    """
    from zenml.integrations.vllm.flavors import VLLMModelDeployerFlavor

    return [VLLMModelDeployerFlavor]

Functions

get_logger(logger_name: str) -> logging.Logger

Main function to get logger name,.

Parameters:

Name	Type	Description	Default
logger_name	str	Name of logger to initialize.	required

Returns:

Type	Description
Logger	A logger object.

Source code in src/zenml/logger.py

def get_logger(logger_name: str) -> logging.Logger:
    """Main function to get logger name,.

    Args:
        logger_name: Name of logger to initialize.

    Returns:
        A logger object.
    """
    logger = logging.getLogger(logger_name)
    logger.setLevel(get_logging_level().value)
    logger.addHandler(get_console_handler())

    logger.propagate = False
    return logger

Modules

flavors

vLLM integration flavors.

Classes

VLLMModelDeployerConfig(warn_about_plain_text_secrets: bool = False, **kwargs: Any)

Bases: BaseModelDeployerConfig

Configuration for vLLM Inference model deployer.

Source code in src/zenml/stack/stack_component.py

def __init__(
    self, warn_about_plain_text_secrets: bool = False, **kwargs: Any
) -> None:
    """Ensures that secret references don't clash with pydantic validation.

    StackComponents allow the specification of all their string attributes
    using secret references of the form `{{secret_name.key}}`. This however
    is only possible when the stack component does not perform any explicit
    validation of this attribute using pydantic validators. If this were
    the case, the validation would run on the secret reference and would
    fail or in the worst case, modify the secret reference and lead to
    unexpected behavior. This method ensures that no attributes that require
    custom pydantic validation are set as secret references.

    Args:
        warn_about_plain_text_secrets: If true, then warns about using
            plain-text secrets.
        **kwargs: Arguments to initialize this stack component.

    Raises:
        ValueError: If an attribute that requires custom pydantic validation
            is passed as a secret reference, or if the `name` attribute
            was passed as a secret reference.
    """
    for key, value in kwargs.items():
        try:
            field = self.__class__.model_fields[key]
        except KeyError:
            # Value for a private attribute or non-existing field, this
            # will fail during the upcoming pydantic validation
            continue

        if value is None:
            continue

        if not secret_utils.is_secret_reference(value):
            if (
                secret_utils.is_secret_field(field)
                and warn_about_plain_text_secrets
            ):
                logger.warning(
                    "You specified a plain-text value for the sensitive "
                    f"attribute `{key}` for a `{self.__class__.__name__}` "
                    "stack component. This is currently only a warning, "
                    "but future versions of ZenML will require you to pass "
                    "in sensitive information as secrets. Check out the "
                    "documentation on how to configure your stack "
                    "components with secrets here: "
                    "https://docs.zenml.io/getting-started/deploying-zenml/secret-management"
                )
            continue

        if pydantic_utils.has_validators(
            pydantic_class=self.__class__, field_name=key
        ):
            raise ValueError(
                f"Passing the stack component attribute `{key}` as a "
                "secret reference is not allowed as additional validation "
                "is required for this attribute."
            )

    super().__init__(**kwargs)

VLLMModelDeployerFlavor

Bases: BaseModelDeployerFlavor

vLLM model deployer flavor.

Attributes

config_class: Type[VLLMModelDeployerConfig] property

Returns VLLMModelDeployerConfig config class.

Returns:

Type	Description
Type[VLLMModelDeployerConfig]	The config class.

docs_url: Optional[str] property

A url to point at docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor docs url.

implementation_class: Type[VLLMModelDeployer] property

Implementation class for this flavor.

Returns:

Type	Description
Type[VLLMModelDeployer]	The implementation class.

logo_url: str property

A url to represent the flavor in the dashboard.

Returns:

Type	Description
str	The flavor logo.

name: str property

Name of the flavor.

Returns:

Type	Description
str	The name of the flavor.

sdk_docs_url: Optional[str] property

A url to point at SDK docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor SDK docs url.

Modules

vllm_model_deployer_flavor

vLLM model deployer flavor.

Classes

VLLMModelDeployerConfig(warn_about_plain_text_secrets: bool = False, **kwargs: Any)

Bases: BaseModelDeployerConfig

Configuration for vLLM Inference model deployer.

Source code in src/zenml/stack/stack_component.py

def __init__(
    self, warn_about_plain_text_secrets: bool = False, **kwargs: Any
) -> None:
    """Ensures that secret references don't clash with pydantic validation.

    StackComponents allow the specification of all their string attributes
    using secret references of the form `{{secret_name.key}}`. This however
    is only possible when the stack component does not perform any explicit
    validation of this attribute using pydantic validators. If this were
    the case, the validation would run on the secret reference and would
    fail or in the worst case, modify the secret reference and lead to
    unexpected behavior. This method ensures that no attributes that require
    custom pydantic validation are set as secret references.

    Args:
        warn_about_plain_text_secrets: If true, then warns about using
            plain-text secrets.
        **kwargs: Arguments to initialize this stack component.

    Raises:
        ValueError: If an attribute that requires custom pydantic validation
            is passed as a secret reference, or if the `name` attribute
            was passed as a secret reference.
    """
    for key, value in kwargs.items():
        try:
            field = self.__class__.model_fields[key]
        except KeyError:
            # Value for a private attribute or non-existing field, this
            # will fail during the upcoming pydantic validation
            continue

        if value is None:
            continue

        if not secret_utils.is_secret_reference(value):
            if (
                secret_utils.is_secret_field(field)
                and warn_about_plain_text_secrets
            ):
                logger.warning(
                    "You specified a plain-text value for the sensitive "
                    f"attribute `{key}` for a `{self.__class__.__name__}` "
                    "stack component. This is currently only a warning, "
                    "but future versions of ZenML will require you to pass "
                    "in sensitive information as secrets. Check out the "
                    "documentation on how to configure your stack "
                    "components with secrets here: "
                    "https://docs.zenml.io/getting-started/deploying-zenml/secret-management"
                )
            continue

        if pydantic_utils.has_validators(
            pydantic_class=self.__class__, field_name=key
        ):
            raise ValueError(
                f"Passing the stack component attribute `{key}` as a "
                "secret reference is not allowed as additional validation "
                "is required for this attribute."
            )

    super().__init__(**kwargs)

VLLMModelDeployerFlavor

Bases: BaseModelDeployerFlavor

vLLM model deployer flavor.

Attributes

config_class: Type[VLLMModelDeployerConfig] property

Returns VLLMModelDeployerConfig config class.

Returns:

Type	Description
Type[VLLMModelDeployerConfig]	The config class.

docs_url: Optional[str] property

A url to point at docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor docs url.

implementation_class: Type[VLLMModelDeployer] property

Implementation class for this flavor.

Returns:

Type	Description
Type[VLLMModelDeployer]	The implementation class.

logo_url: str property

A url to represent the flavor in the dashboard.

Returns:

Type	Description
str	The flavor logo.

name: str property

Name of the flavor.

Returns:

Type	Description
str	The name of the flavor.

sdk_docs_url: Optional[str] property

A url to point at SDK docs explaining this flavor.

Returns:

Type	Description
Optional[str]	A flavor SDK docs url.

model_deployers

Initialization of the vLLM model deployers.

Classes

VLLMModelDeployer(name: str, id: UUID, config: StackComponentConfig, flavor: str, type: StackComponentType, user: Optional[UUID], created: datetime, updated: datetime, labels: Optional[Dict[str, Any]] = None, connector_requirements: Optional[ServiceConnectorRequirements] = None, connector: Optional[UUID] = None, connector_resource_id: Optional[str] = None, *args: Any, **kwargs: Any)

Bases: BaseModelDeployer

vLLM Inference Server.

Source code in src/zenml/stack/stack_component.py

def __init__(
    self,
    name: str,
    id: UUID,
    config: StackComponentConfig,
    flavor: str,
    type: StackComponentType,
    user: Optional[UUID],
    created: datetime,
    updated: datetime,
    labels: Optional[Dict[str, Any]] = None,
    connector_requirements: Optional[ServiceConnectorRequirements] = None,
    connector: Optional[UUID] = None,
    connector_resource_id: Optional[str] = None,
    *args: Any,
    **kwargs: Any,
):
    """Initializes a StackComponent.

    Args:
        name: The name of the component.
        id: The unique ID of the component.
        config: The config of the component.
        flavor: The flavor of the component.
        type: The type of the component.
        user: The ID of the user who created the component.
        created: The creation time of the component.
        updated: The last update time of the component.
        labels: The labels of the component.
        connector_requirements: The requirements for the connector.
        connector: The ID of a connector linked to the component.
        connector_resource_id: The custom resource ID to access through
            the connector.
        *args: Additional positional arguments.
        **kwargs: Additional keyword arguments.

    Raises:
        ValueError: If a secret reference is passed as name.
    """
    if secret_utils.is_secret_reference(name):
        raise ValueError(
            "Passing the `name` attribute of a stack component as a "
            "secret reference is not allowed."
        )

    self.id = id
    self.name = name
    self._config = config
    self.flavor = flavor
    self.type = type
    self.user = user
    self.created = created
    self.updated = updated
    self.labels = labels
    self.connector_requirements = connector_requirements
    self.connector = connector
    self.connector_resource_id = connector_resource_id
    self._connector_instance: Optional[ServiceConnector] = None

Attributes

config: VLLMModelDeployerConfig property

Returns the VLLMModelDeployerConfig config.

Returns:

Type	Description
VLLMModelDeployerConfig	The configuration.

local_path: str property

Returns the path to the root directory.

This is where all configurations for vLLM deployment daemon processes are stored.

If the service path is not set in the config by the user, the path is set to a local default path according to the component ID.

Returns:

Type	Description
str	The path to the local service root directory.

Functions

get_model_server_info(service_instance: VLLMDeploymentService) -> Dict[str, Optional[str]] staticmethod

Return implementation specific information on the model server.

Parameters:

Name	Type	Description	Default
service_instance	VLLMDeploymentService	vLLM deployment service object	required

Returns:

Type	Description
Dict[str, Optional[str]]	A dictionary containing the model server information.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

@staticmethod
def get_model_server_info(  # type: ignore[override]
    service_instance: "VLLMDeploymentService",
) -> Dict[str, Optional[str]]:
    """Return implementation specific information on the model server.

    Args:
        service_instance: vLLM deployment service object

    Returns:
        A dictionary containing the model server information.
    """
    return {
        "HEALTH_CHECK_URL": service_instance.get_healthcheck_url(),
        "PREDICTION_URL": service_instance.get_prediction_url(),
        "SERVICE_PATH": service_instance.status.runtime_path,
        "DAEMON_PID": str(service_instance.status.pid),
    }

get_service_path(id_: UUID) -> str staticmethod

Get the path where local vLLM service information is stored.

This includes the deployment service configuration, PID and log files are stored.

Parameters:

Name	Type	Description	Default
id_	UUID	The ID of the vLLM model deployer.	required

Returns:

Type	Description
str	The service path.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

@staticmethod
def get_service_path(id_: UUID) -> str:
    """Get the path where local vLLM service information is stored.

    This includes the deployment service configuration, PID and log files
    are stored.

    Args:
        id_: The ID of the vLLM model deployer.

    Returns:
        The service path.
    """
    service_path = os.path.join(
        GlobalConfiguration().local_stores_path,
        str(id_),
    )
    create_dir_recursive_if_not_exists(service_path)
    return service_path

perform_delete_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT, force: bool = False) -> None

Method to delete all configuration of a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to delete.	required
timeout	int	Timeout in seconds to wait for the service to stop.	DEFAULT_SERVICE_START_STOP_TIMEOUT
force	bool	If True, force the service to stop.	False

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_delete_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
    force: bool = False,
) -> None:
    """Method to delete all configuration of a model server.

    Args:
        service: The service to delete.
        timeout: Timeout in seconds to wait for the service to stop.
        force: If True, force the service to stop.
    """
    service = cast(VLLMDeploymentService, service)
    self._clean_up_existing_service(
        existing_service=service, timeout=timeout, force=force
    )

perform_deploy_model(id: UUID, config: ServiceConfig, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT) -> BaseService

Create a new vLLM deployment service or update an existing one.

This should serve the supplied model and deployment configuration.

This method has two modes of operation, depending on the replace argument value:

• if replace is False, calling this method will create a new vLLM deployment server to reflect the model and other configuration parameters specified in the supplied vLLM service config.

• if replace is True, this method will first attempt to find an existing vLLM deployment service that is equivalent to the supplied configuration parameters. Two or more vLLM deployment services are considered equivalent if they have the same pipeline_name, pipeline_step_name and model_name configuration parameters. To put it differently, two vLLM deployment services are equivalent if they serve versions of the same model deployed by the same pipeline step. If an equivalent vLLM deployment is found, it will be updated in place to reflect the new configuration parameters.

Callers should set replace to True if they want a continuous model deployment workflow that doesn't spin up a new vLLM deployment server for each new model version. If multiple equivalent vLLM deployment servers are found, one is selected at random to be updated and the others are deleted.

Parameters:

Name	Type	Description	Default
id	UUID	the UUID of the vLLM model deployer.	required
config	ServiceConfig	the configuration of the model to be deployed with vLLM.	required
timeout	int	the timeout in seconds to wait for the vLLM server to be provisioned and successfully started or updated. If set to 0, the method will return immediately after the vLLM server is provisioned, without waiting for it to fully start.	DEFAULT_SERVICE_START_STOP_TIMEOUT

Returns:

Type	Description
BaseService	The ZenML vLLM deployment service object that can be used to
BaseService	interact with the vLLM model http server.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_deploy_model(
    self,
    id: UUID,
    config: ServiceConfig,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
) -> BaseService:
    """Create a new vLLM deployment service or update an existing one.

    This should serve the supplied model and deployment configuration.

    This method has two modes of operation, depending on the `replace`
    argument value:

      * if `replace` is False, calling this method will create a new vLLM
        deployment server to reflect the model and other configuration
        parameters specified in the supplied vLLM service `config`.

      * if `replace` is True, this method will first attempt to find an
        existing vLLM deployment service that is *equivalent* to the
        supplied configuration parameters. Two or more vLLM deployment
        services are considered equivalent if they have the same
        `pipeline_name`, `pipeline_step_name` and `model_name` configuration
        parameters. To put it differently, two vLLM deployment services
        are equivalent if they serve versions of the same model deployed by
        the same pipeline step. If an equivalent vLLM deployment is found,
        it will be updated in place to reflect the new configuration
        parameters.

    Callers should set `replace` to True if they want a continuous model
    deployment workflow that doesn't spin up a new vLLM deployment
    server for each new model version. If multiple equivalent vLLM
    deployment servers are found, one is selected at random to be updated
    and the others are deleted.

    Args:
        id: the UUID of the vLLM model deployer.
        config: the configuration of the model to be deployed with vLLM.
        timeout: the timeout in seconds to wait for the vLLM server
            to be provisioned and successfully started or updated. If set
            to 0, the method will return immediately after the vLLM
            server is provisioned, without waiting for it to fully start.

    Returns:
        The ZenML vLLM deployment service object that can be used to
        interact with the vLLM model http server.
    """
    config = cast(VLLMServiceConfig, config)
    service = self._create_new_service(
        id=id, timeout=timeout, config=config
    )
    logger.info(f"Created a new vLLM deployment service: {service}")
    return service

perform_start_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT) -> BaseService

Method to start a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to start.	required
timeout	int	Timeout in seconds to wait for the service to start.	DEFAULT_SERVICE_START_STOP_TIMEOUT

Returns:

Type	Description
BaseService	The started service.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_start_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
) -> BaseService:
    """Method to start a model server.

    Args:
        service: The service to start.
        timeout: Timeout in seconds to wait for the service to start.

    Returns:
        The started service.
    """
    service.start(timeout=timeout)
    return service

perform_stop_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT, force: bool = False) -> BaseService

Method to stop a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to stop.	required
timeout	int	Timeout in seconds to wait for the service to stop.	DEFAULT_SERVICE_START_STOP_TIMEOUT
force	bool	If True, force the service to stop.	False

Returns:

Type	Description
BaseService	The stopped service.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_stop_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
    force: bool = False,
) -> BaseService:
    """Method to stop a model server.

    Args:
        service: The service to stop.
        timeout: Timeout in seconds to wait for the service to stop.
        force: If True, force the service to stop.

    Returns:
        The stopped service.
    """
    service.stop(timeout=timeout, force=force)
    return service

Modules

vllm_model_deployer

Implementation of the vLLM Model Deployer.

Classes

VLLMModelDeployer(name: str, id: UUID, config: StackComponentConfig, flavor: str, type: StackComponentType, user: Optional[UUID], created: datetime, updated: datetime, labels: Optional[Dict[str, Any]] = None, connector_requirements: Optional[ServiceConnectorRequirements] = None, connector: Optional[UUID] = None, connector_resource_id: Optional[str] = None, *args: Any, **kwargs: Any)

Bases: BaseModelDeployer

vLLM Inference Server.

Source code in src/zenml/stack/stack_component.py

def __init__(
    self,
    name: str,
    id: UUID,
    config: StackComponentConfig,
    flavor: str,
    type: StackComponentType,
    user: Optional[UUID],
    created: datetime,
    updated: datetime,
    labels: Optional[Dict[str, Any]] = None,
    connector_requirements: Optional[ServiceConnectorRequirements] = None,
    connector: Optional[UUID] = None,
    connector_resource_id: Optional[str] = None,
    *args: Any,
    **kwargs: Any,
):
    """Initializes a StackComponent.

    Args:
        name: The name of the component.
        id: The unique ID of the component.
        config: The config of the component.
        flavor: The flavor of the component.
        type: The type of the component.
        user: The ID of the user who created the component.
        created: The creation time of the component.
        updated: The last update time of the component.
        labels: The labels of the component.
        connector_requirements: The requirements for the connector.
        connector: The ID of a connector linked to the component.
        connector_resource_id: The custom resource ID to access through
            the connector.
        *args: Additional positional arguments.
        **kwargs: Additional keyword arguments.

    Raises:
        ValueError: If a secret reference is passed as name.
    """
    if secret_utils.is_secret_reference(name):
        raise ValueError(
            "Passing the `name` attribute of a stack component as a "
            "secret reference is not allowed."
        )

    self.id = id
    self.name = name
    self._config = config
    self.flavor = flavor
    self.type = type
    self.user = user
    self.created = created
    self.updated = updated
    self.labels = labels
    self.connector_requirements = connector_requirements
    self.connector = connector
    self.connector_resource_id = connector_resource_id
    self._connector_instance: Optional[ServiceConnector] = None

Attributes

config: VLLMModelDeployerConfig property

Returns the VLLMModelDeployerConfig config.

Returns:

Type	Description
VLLMModelDeployerConfig	The configuration.

local_path: str property

Returns the path to the root directory.

This is where all configurations for vLLM deployment daemon processes are stored.

If the service path is not set in the config by the user, the path is set to a local default path according to the component ID.

Returns:

Type	Description
str	The path to the local service root directory.

Functions

get_model_server_info(service_instance: VLLMDeploymentService) -> Dict[str, Optional[str]] staticmethod

Return implementation specific information on the model server.

Parameters:

Name	Type	Description	Default
service_instance	VLLMDeploymentService	vLLM deployment service object	required

Returns:

Type	Description
Dict[str, Optional[str]]	A dictionary containing the model server information.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

@staticmethod
def get_model_server_info(  # type: ignore[override]
    service_instance: "VLLMDeploymentService",
) -> Dict[str, Optional[str]]:
    """Return implementation specific information on the model server.

    Args:
        service_instance: vLLM deployment service object

    Returns:
        A dictionary containing the model server information.
    """
    return {
        "HEALTH_CHECK_URL": service_instance.get_healthcheck_url(),
        "PREDICTION_URL": service_instance.get_prediction_url(),
        "SERVICE_PATH": service_instance.status.runtime_path,
        "DAEMON_PID": str(service_instance.status.pid),
    }

get_service_path(id_: UUID) -> str staticmethod

Get the path where local vLLM service information is stored.

This includes the deployment service configuration, PID and log files are stored.

Parameters:

Name	Type	Description	Default
id_	UUID	The ID of the vLLM model deployer.	required

Returns:

Type	Description
str	The service path.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

@staticmethod
def get_service_path(id_: UUID) -> str:
    """Get the path where local vLLM service information is stored.

    This includes the deployment service configuration, PID and log files
    are stored.

    Args:
        id_: The ID of the vLLM model deployer.

    Returns:
        The service path.
    """
    service_path = os.path.join(
        GlobalConfiguration().local_stores_path,
        str(id_),
    )
    create_dir_recursive_if_not_exists(service_path)
    return service_path

perform_delete_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT, force: bool = False) -> None

Method to delete all configuration of a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to delete.	required
timeout	int	Timeout in seconds to wait for the service to stop.	DEFAULT_SERVICE_START_STOP_TIMEOUT
force	bool	If True, force the service to stop.	False

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_delete_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
    force: bool = False,
) -> None:
    """Method to delete all configuration of a model server.

    Args:
        service: The service to delete.
        timeout: Timeout in seconds to wait for the service to stop.
        force: If True, force the service to stop.
    """
    service = cast(VLLMDeploymentService, service)
    self._clean_up_existing_service(
        existing_service=service, timeout=timeout, force=force
    )

perform_deploy_model(id: UUID, config: ServiceConfig, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT) -> BaseService

Create a new vLLM deployment service or update an existing one.

This should serve the supplied model and deployment configuration.

This method has two modes of operation, depending on the replace argument value:

• if replace is False, calling this method will create a new vLLM deployment server to reflect the model and other configuration parameters specified in the supplied vLLM service config.

• if replace is True, this method will first attempt to find an existing vLLM deployment service that is equivalent to the supplied configuration parameters. Two or more vLLM deployment services are considered equivalent if they have the same pipeline_name, pipeline_step_name and model_name configuration parameters. To put it differently, two vLLM deployment services are equivalent if they serve versions of the same model deployed by the same pipeline step. If an equivalent vLLM deployment is found, it will be updated in place to reflect the new configuration parameters.

Callers should set replace to True if they want a continuous model deployment workflow that doesn't spin up a new vLLM deployment server for each new model version. If multiple equivalent vLLM deployment servers are found, one is selected at random to be updated and the others are deleted.

Parameters:

Name	Type	Description	Default
id	UUID	the UUID of the vLLM model deployer.	required
config	ServiceConfig	the configuration of the model to be deployed with vLLM.	required
timeout	int	the timeout in seconds to wait for the vLLM server to be provisioned and successfully started or updated. If set to 0, the method will return immediately after the vLLM server is provisioned, without waiting for it to fully start.	DEFAULT_SERVICE_START_STOP_TIMEOUT

Returns:

Type	Description
BaseService	The ZenML vLLM deployment service object that can be used to
BaseService	interact with the vLLM model http server.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_deploy_model(
    self,
    id: UUID,
    config: ServiceConfig,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
) -> BaseService:
    """Create a new vLLM deployment service or update an existing one.

    This should serve the supplied model and deployment configuration.

    This method has two modes of operation, depending on the `replace`
    argument value:

      * if `replace` is False, calling this method will create a new vLLM
        deployment server to reflect the model and other configuration
        parameters specified in the supplied vLLM service `config`.

      * if `replace` is True, this method will first attempt to find an
        existing vLLM deployment service that is *equivalent* to the
        supplied configuration parameters. Two or more vLLM deployment
        services are considered equivalent if they have the same
        `pipeline_name`, `pipeline_step_name` and `model_name` configuration
        parameters. To put it differently, two vLLM deployment services
        are equivalent if they serve versions of the same model deployed by
        the same pipeline step. If an equivalent vLLM deployment is found,
        it will be updated in place to reflect the new configuration
        parameters.

    Callers should set `replace` to True if they want a continuous model
    deployment workflow that doesn't spin up a new vLLM deployment
    server for each new model version. If multiple equivalent vLLM
    deployment servers are found, one is selected at random to be updated
    and the others are deleted.

    Args:
        id: the UUID of the vLLM model deployer.
        config: the configuration of the model to be deployed with vLLM.
        timeout: the timeout in seconds to wait for the vLLM server
            to be provisioned and successfully started or updated. If set
            to 0, the method will return immediately after the vLLM
            server is provisioned, without waiting for it to fully start.

    Returns:
        The ZenML vLLM deployment service object that can be used to
        interact with the vLLM model http server.
    """
    config = cast(VLLMServiceConfig, config)
    service = self._create_new_service(
        id=id, timeout=timeout, config=config
    )
    logger.info(f"Created a new vLLM deployment service: {service}")
    return service

perform_start_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT) -> BaseService

Method to start a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to start.	required
timeout	int	Timeout in seconds to wait for the service to start.	DEFAULT_SERVICE_START_STOP_TIMEOUT

Returns:

Type	Description
BaseService	The started service.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_start_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
) -> BaseService:
    """Method to start a model server.

    Args:
        service: The service to start.
        timeout: Timeout in seconds to wait for the service to start.

    Returns:
        The started service.
    """
    service.start(timeout=timeout)
    return service

perform_stop_model(service: BaseService, timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT, force: bool = False) -> BaseService

Method to stop a model server.

Parameters:

Name	Type	Description	Default
service	BaseService	The service to stop.	required
timeout	int	Timeout in seconds to wait for the service to stop.	DEFAULT_SERVICE_START_STOP_TIMEOUT
force	bool	If True, force the service to stop.	False

Returns:

Type	Description
BaseService	The stopped service.

Source code in src/zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

def perform_stop_model(
    self,
    service: BaseService,
    timeout: int = DEFAULT_SERVICE_START_STOP_TIMEOUT,
    force: bool = False,
) -> BaseService:
    """Method to stop a model server.

    Args:
        service: The service to stop.
        timeout: Timeout in seconds to wait for the service to stop.
        force: If True, force the service to stop.

    Returns:
        The stopped service.
    """
    service.stop(timeout=timeout, force=force)
    return service

Functions

services

Initialization of the vLLM Inference Server.

Classes

Modules

vllm_deployment

Implementation of the vLLM Inference Server Service.

Classes

VLLMDeploymentEndpoint(*args: Any, **kwargs: Any)

Bases: LocalDaemonServiceEndpoint

A service endpoint exposed by the vLLM deployment daemon.

Attributes:

Name	Type	Description
config	VLLMDeploymentEndpointConfig	service endpoint configuration

Source code in src/zenml/services/service_endpoint.py

def __init__(
    self,
    *args: Any,
    **kwargs: Any,
) -> None:
    """Initialize the service endpoint.

    Args:
        *args: positional arguments.
        **kwargs: keyword arguments.
    """
    super().__init__(*args, **kwargs)
    self.config.name = self.config.name or self.__class__.__name__

Attributes

prediction_url: Optional[str] property

Gets the prediction URL for the endpoint.

Returns:

Type	Description
Optional[str]	the prediction URL for the endpoint

VLLMDeploymentEndpointConfig

Bases: LocalDaemonServiceEndpointConfig

vLLM deployment service configuration.

Attributes:

Name	Type	Description
prediction_url_path	str	URI subpath for prediction requests

VLLMDeploymentService(config: VLLMServiceConfig, **attrs: Any)

Bases: LocalDaemonService, BaseDeploymentService

vLLM Inference Server Deployment Service.

Initialize the vLLM deployment service.

Parameters:

Name	Type	Description	Default
config	VLLMServiceConfig	service configuration	required
attrs	Any	additional attributes to set on the service	{}

Source code in src/zenml/integrations/vllm/services/vllm_deployment.py

def __init__(self, config: VLLMServiceConfig, **attrs: Any):
    """Initialize the vLLM deployment service.

    Args:
        config: service configuration
        attrs: additional attributes to set on the service
    """
    if isinstance(config, VLLMServiceConfig) and "endpoint" not in attrs:
        endpoint = VLLMDeploymentEndpoint(
            config=VLLMDeploymentEndpointConfig(
                protocol=ServiceEndpointProtocol.HTTP,
                port=config.port,
                ip_address=config.host or DEFAULT_LOCAL_SERVICE_IP_ADDRESS,
                prediction_url_path=VLLM_PREDICTION_URL_PATH,
            ),
            monitor=HTTPEndpointHealthMonitor(
                config=HTTPEndpointHealthMonitorConfig(
                    healthcheck_uri_path=VLLM_HEALTHCHECK_URL_PATH,
                )
            ),
        )
        attrs["endpoint"] = endpoint
    super().__init__(config=config, **attrs)

Attributes

prediction_url: Optional[str] property

Gets the prediction URL for the endpoint.

Returns:

Type	Description
Optional[str]	the prediction URL for the endpoint

Functions

predict(data: Any) -> Any

Make a prediction using the service.

Parameters:

Name	Type	Description	Default
data	Any	data to make a prediction on	required

Returns:

Type	Description
Any	The prediction result.

Raises:

Type	Description
Exception	if the service is not running
ValueError	if the prediction endpoint is unknown.

Source code in src/zenml/integrations/vllm/services/vllm_deployment.py

def predict(self, data: "Any") -> "Any":
    """Make a prediction using the service.

    Args:
        data: data to make a prediction on

    Returns:
        The prediction result.

    Raises:
        Exception: if the service is not running
        ValueError: if the prediction endpoint is unknown.
    """
    if not self.is_running:
        raise Exception(
            "vLLM Inference service is not running. "
            "Please start the service before making predictions."
        )
    if self.endpoint.prediction_url is not None:
        from openai import OpenAI

        client = OpenAI(
            api_key="EMPTY",
            base_url=self.endpoint.prediction_url,
        )
        models = client.models.list()
        model = models.data[0].id
        result = client.completions.create(model=model, prompt=data)
        # TODO: We can add support for client.chat.completions.create
    else:
        raise ValueError("No endpoint known for prediction.")
    return result

run() -> None

Start the service.

Source code in src/zenml/integrations/vllm/services/vllm_deployment.py

def run(self) -> None:
    """Start the service."""
    logger.info(
        "Starting vLLM inference server service as blocking "
        "process... press CTRL+C once to stop it."
    )

    self.endpoint.prepare_for_start()

    import uvloop
    from vllm.entrypoints.openai.api_server import (
        run_server,
    )
    from vllm.entrypoints.openai.cli_args import (
        make_arg_parser,
    )
    from vllm.utils import (
        FlexibleArgumentParser,
    )

    try:
        parser: argparse.ArgumentParser = make_arg_parser(
            FlexibleArgumentParser()  # type: ignore[no-untyped-call]
        )
        # pass in empty list to get default args
        # otherwise it will try to get the args from sys.argv
        # and if there's a --config in there, it will want to use
        # that file for vLLM configuration, which is not what we want
        args: argparse.Namespace = parser.parse_args(args=[])
        # Override port with the available port
        self.config.port = self.endpoint.status.port or self.config.port

        # Update the arguments in place
        args.__dict__.update(self.config.model_dump())
        uvloop.run(run_server(args=args))
    except KeyboardInterrupt:
        logger.info("Stopping vLLM prediction service...")

VLLMServiceConfig(**data: Any)

Bases: LocalDaemonServiceConfig

vLLM service configurations.

Source code in src/zenml/services/service.py

def __init__(self, **data: Any):
    """Initialize the service configuration.

    Args:
        **data: keyword arguments.

    Raises:
        ValueError: if neither 'name' nor 'model_name' is set.
    """
    super().__init__(**data)
    if self.name or self.model_name:
        self.service_name = data.get(
            "service_name",
            f"{ZENM_ENDPOINT_PREFIX}{self.name or self.model_name}",
        )
    else:
        raise ValueError("Either 'name' or 'model_name' must be set.")

Functions