zenml-nightly 0.68.1.dev20241102__py3-none-any.whl → 0.68.1.dev20241106__py3-none-any.whl

zenml/integrations/vllm/model_deployers/vllm_model_deployer.py

@@ -0,0 +1,263 @@
+#  Copyright (c) ZenML GmbH 2024. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at:
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+"""Implementation of the vLLM Model Deployer."""
+
+import os
+import shutil
+from typing import ClassVar, Dict, Optional, Type, cast
+from uuid import UUID
+
+from zenml.config.global_config import GlobalConfiguration
+from zenml.constants import DEFAULT_SERVICE_START_STOP_TIMEOUT
+from zenml.integrations.vllm.flavors.vllm_model_deployer_flavor import (
+    VLLMModelDeployerConfig,
+    VLLMModelDeployerFlavor,
+)
+from zenml.integrations.vllm.services.vllm_deployment import (
+    VLLMDeploymentService,
+    VLLMServiceConfig,
+)
+from zenml.logger import get_logger
+from zenml.model_deployers import BaseModelDeployer, BaseModelDeployerFlavor
+from zenml.services.service import BaseService, ServiceConfig
+from zenml.utils.io_utils import create_dir_recursive_if_not_exists
+
+logger = get_logger(__name__)
+
+
+class VLLMModelDeployer(BaseModelDeployer):
+    """vLLM Inference Server."""
+
+    NAME: ClassVar[str] = "VLLM"
+    FLAVOR: ClassVar[Type[BaseModelDeployerFlavor]] = VLLMModelDeployerFlavor
+
+    _service_path: Optional[str] = None
+
+    @property
+    def config(self) -> VLLMModelDeployerConfig:
+        """Returns the `VLLMModelDeployerConfig` config.
+
+        Returns:
+            The configuration.
+        """
+        return cast(VLLMModelDeployerConfig, self._config)
+
+    @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
+
+    @property
+    def local_path(self) -> str:
+        """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:
+            The path to the local service root directory.
+        """
+        if self._service_path is not None:
+            return self._service_path
+
+        if self.config.service_path:
+            self._service_path = self.config.service_path
+        else:
+            self._service_path = self.get_service_path(self.id)
+
+        create_dir_recursive_if_not_exists(self._service_path)
+        return self._service_path
+
+    @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),
+        }
+
+    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
+
+    def _clean_up_existing_service(
+        self,
+        timeout: int,
+        force: bool,
+        existing_service: VLLMDeploymentService,
+    ) -> None:
+        # stop the older service
+        existing_service.stop(timeout=timeout, force=force)
+
+        # delete the old configuration file
+        if existing_service.status.runtime_path:
+            shutil.rmtree(existing_service.status.runtime_path)
+
+    # the step will receive a config from the user that mentions the number
+    # of workers etc.the step implementation will create a new config using
+    # all values from the user and add values like pipeline name, model_uri
+    def _create_new_service(
+        self, id: UUID, timeout: int, config: VLLMServiceConfig
+    ) -> VLLMDeploymentService:
+        """Creates a new VLLMDeploymentService.
+
+        Args:
+            id: the ID of the vLLM deployment service to be created or updated.
+            timeout: the timeout in seconds to wait for the vLLM server
+                to be provisioned and successfully started or updated.
+            config: the configuration of the model to be deployed with vLLM.
+
+        Returns:
+            The VLLMDeploymentService object that can be used to interact
+            with the vLLM model server.
+        """
+        # set the root runtime path with the stack component's UUID
+        config.root_runtime_path = self.local_path
+        # create a new service for the new model
+        service = VLLMDeploymentService(uuid=id, config=config)
+        service.start(timeout=timeout)
+
+        return service
+
+    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
+
+    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
+
+    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
+        )

zenml/integrations/vllm/services/__init__.py

@@ -0,0 +1,19 @@
+#  Copyright (c) ZenML GmbH 2024. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at:
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+"""Initialization of the vLLM Inference Server."""
+
+from zenml.integrations.vllm.services.vllm_deployment import (  # noqa
+    VLLMDeploymentService,
+    VLLMServiceConfig,
+)

zenml/integrations/vllm/services/vllm_deployment.py

@@ -0,0 +1,197 @@
+#  Copyright (c) ZenML GmbH 2024. All Rights Reserved.
+#
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at:
+#
+#       https://www.apache.org/licenses/LICENSE-2.0
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
+#  or implied. See the License for the specific language governing
+#  permissions and limitations under the License.
+"""Implementation of the vLLM Inference Server Service."""
+
+import os
+from typing import Any, List, Optional, Union
+
+from zenml.constants import DEFAULT_LOCAL_SERVICE_IP_ADDRESS
+from zenml.logger import get_logger
+from zenml.services import (
+    HTTPEndpointHealthMonitor,
+    HTTPEndpointHealthMonitorConfig,
+    LocalDaemonService,
+    LocalDaemonServiceConfig,
+    LocalDaemonServiceEndpoint,
+    LocalDaemonServiceEndpointConfig,
+    ServiceEndpointProtocol,
+    ServiceType,
+)
+from zenml.services.service import BaseDeploymentService
+
+logger = get_logger(__name__)
+
+
+VLLM_PREDICTION_URL_PATH = "v1"
+VLLM_HEALTHCHECK_URL_PATH = "health"
+
+
+class VLLMDeploymentEndpointConfig(LocalDaemonServiceEndpointConfig):
+    """vLLM deployment service configuration.
+
+    Attributes:
+        prediction_url_path: URI subpath for prediction requests
+    """
+
+    prediction_url_path: str
+
+
+class VLLMDeploymentEndpoint(LocalDaemonServiceEndpoint):
+    """A service endpoint exposed by the vLLM deployment daemon.
+
+    Attributes:
+        config: service endpoint configuration
+    """
+
+    config: VLLMDeploymentEndpointConfig
+    monitor: HTTPEndpointHealthMonitor
+
+    @property
+    def prediction_url(self) -> Optional[str]:
+        """Gets the prediction URL for the endpoint.
+
+        Returns:
+            the prediction URL for the endpoint
+        """
+        uri = self.status.uri
+        if not uri:
+            return None
+        return os.path.join(uri, self.config.prediction_url_path)
+
+
+class VLLMServiceConfig(LocalDaemonServiceConfig):
+    """vLLM service configurations."""
+
+    model: str
+    port: int
+    host: Optional[str] = None
+    blocking: bool = True
+    # If unspecified, model name or path will be used.
+    tokenizer: Optional[str] = None
+    served_model_name: Optional[Union[str, List[str]]] = None
+    # Trust remote code from huggingface.
+    trust_remote_code: Optional[bool] = False
+    # ['auto', 'slow', 'mistral']
+    tokenizer_mode: Optional[str] = "auto"
+    # ['auto', 'half', 'float16', 'bfloat16', 'float', 'float32']
+    dtype: Optional[str] = "auto"
+    # The specific model version to use. It can be a branch name, a tag name, or a commit id.
+    # If unspecified, will use the default version.
+    revision: Optional[str] = None
+
+
+class VLLMDeploymentService(LocalDaemonService, BaseDeploymentService):
+    """vLLM Inference Server Deployment Service."""
+
+    SERVICE_TYPE = ServiceType(
+        name="vllm-deployment",
+        type="model-serving",
+        flavor="vllm",
+        description="vLLM Inference prediction service",
+    )
+    config: VLLMServiceConfig
+    endpoint: VLLMDeploymentEndpoint
+
+    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)
+
+    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 = make_arg_parser(FlexibleArgumentParser())
+            args = parser.parse_args()
+            # Override port with the available port
+            self.config.port = self.endpoint.status.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...")
+
+    @property
+    def prediction_url(self) -> Optional[str]:
+        """Gets the prediction URL for the endpoint.
+
+        Returns:
+            the prediction URL for the endpoint
+        """
+        if not self.is_running:
+            return None
+        return self.endpoint.prediction_url_path
+
+    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

zenml/integrations/whylogs/materializers/whylogs_materializer.py

@@ -14,7 +14,6 @@
 """Implementation of the whylogs materializer."""
 
 import os
-import tempfile
 from typing import Any, ClassVar, Dict, Tuple, Type, cast
 
 from whylogs.core import DatasetProfileView  # type: ignore
@@ -51,18 +50,14 @@ class WhylogsMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, PROFILE_FILENAME)
 
-        # Create a temporary folder
-        temp_dir = tempfile.mkdtemp(prefix="zenml-temp-")
-        temp_file = os.path.join(str(temp_dir), PROFILE_FILENAME)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), PROFILE_FILENAME)
 
-        # Copy from artifact store to temporary file
-        fileio.copy(filepath, temp_file)
-        profile_view = DatasetProfileView.read(temp_file)
+            # Copy from artifact store to temporary file
+            fileio.copy(filepath, temp_file)
+            profile_view = DatasetProfileView.read(temp_file)
 
-        # Cleanup and return
-        fileio.rmtree(temp_dir)
-
-        return profile_view
+            return profile_view
 
     def save(self, profile_view: DatasetProfileView) -> None:
         """Writes a whylogs dataset profile view.
@@ -72,15 +67,13 @@ class WhylogsMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, PROFILE_FILENAME)
 
-        # Create a temporary folder
-        temp_dir = tempfile.mkdtemp(prefix="zenml-temp-")
-        temp_file = os.path.join(str(temp_dir), PROFILE_FILENAME)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), PROFILE_FILENAME)
 
-        profile_view.write(temp_file)
+            profile_view.write(temp_file)
 
-        # Copy it into artifact store
-        fileio.copy(temp_file, filepath)
-        fileio.rmtree(temp_dir)
+            # Copy it into artifact store
+            fileio.copy(temp_file, filepath)
 
         try:
             self._upload_to_whylabs(profile_view)

zenml/integrations/xgboost/materializers/xgboost_booster_materializer.py

@@ -14,7 +14,6 @@
 """Implementation of an XGBoost booster materializer."""
 
 import os
-import tempfile
 from typing import Any, ClassVar, Tuple, Type
 
 import xgboost as xgb
@@ -43,18 +42,15 @@ class XgboostBoosterMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, DEFAULT_FILENAME)
 
-        # Create a temporary folder
-        temp_dir = tempfile.mkdtemp(prefix="zenml-temp-")
-        temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
 
-        # Copy from artifact store to temporary file
-        fileio.copy(filepath, temp_file)
-        booster = xgb.Booster()
-        booster.load_model(temp_file)
+            # Copy from artifact store to temporary file
+            fileio.copy(filepath, temp_file)
+            booster = xgb.Booster()
+            booster.load_model(temp_file)
 
-        # Cleanup and return
-        fileio.rmtree(temp_dir)
-        return booster
+            return booster
 
     def save(self, booster: xgb.Booster) -> None:
         """Creates a JSON serialization for a xgboost Booster model.
@@ -64,14 +60,7 @@ class XgboostBoosterMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, DEFAULT_FILENAME)
 
-        # Make a temporary phantom artifact
-        with tempfile.NamedTemporaryFile(
-            mode="w", suffix=".json", delete=False
-        ) as f:
-            booster.save_model(f.name)
-            # Copy it into artifact store
-            fileio.copy(f.name, filepath)
-
-        # Close and remove the temporary file
-        f.close()
-        fileio.remove(f.name)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
+            booster.save_model(temp_file)
+            fileio.copy(temp_file, filepath)

zenml/integrations/xgboost/materializers/xgboost_dmatrix_materializer.py

@@ -14,7 +14,6 @@
 """Implementation of the XGBoost dmatrix materializer."""
 
 import os
-import tempfile
 from typing import TYPE_CHECKING, Any, ClassVar, Dict, Tuple, Type
 
 import xgboost as xgb
@@ -46,17 +45,14 @@ class XgboostDMatrixMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, DEFAULT_FILENAME)
 
-        # Create a temporary folder
-        temp_dir = tempfile.mkdtemp(prefix="zenml-temp-")
-        temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
 
-        # Copy from artifact store to temporary file
-        fileio.copy(filepath, temp_file)
-        matrix = xgb.DMatrix(temp_file)
+            # Copy from artifact store to temporary file
+            fileio.copy(filepath, temp_file)
+            matrix = xgb.DMatrix(temp_file)
 
-        # Cleanup and return
-        fileio.rmtree(temp_dir)
-        return matrix
+            return matrix
 
     def save(self, matrix: xgb.DMatrix) -> None:
         """Creates a binary serialization for a xgboost.DMatrix object.
@@ -66,15 +62,10 @@ class XgboostDMatrixMaterializer(BaseMaterializer):
         """
         filepath = os.path.join(self.uri, DEFAULT_FILENAME)
 
-        # Make a temporary phantom artifact
-        with tempfile.NamedTemporaryFile(mode="wb", delete=False) as f:
-            matrix.save_binary(f.name)
-            # Copy it into artifact store
-            fileio.copy(f.name, filepath)
-
-        # Close and remove the temporary file
-        f.close()
-        fileio.remove(f.name)
+        with self.get_temporary_directory(delete_at_exit=True) as temp_dir:
+            temp_file = os.path.join(str(temp_dir), DEFAULT_FILENAME)
+            matrix.save_binary(temp_file)
+            fileio.copy(temp_file, filepath)
 
     def extract_metadata(
         self, dataset: xgb.DMatrix