feat: Add FlexStart option to DeploymentResourcePool.create, Endpoint.deploy, and Model.deploy (preview)

PiperOrigin-RevId: 785601050

Author: vertex-sdk-bot

google/cloud/aiplatform/preview/models.py

@@ -28,6 +28,7 @@
 from google.cloud.aiplatform import models
 from google.cloud.aiplatform import utils
 from google.cloud.aiplatform.utils import _explanation_utils
+from google.cloud.aiplatform.utils import prediction_utils
 from google.cloud.aiplatform.compat.services import (
     deployment_resource_pool_service_client_v1beta1,
     endpoint_service_client,
@@ -140,6 +141,7 @@ def create(
         create_request_timeout: Optional[float] = None,
         required_replica_count: Optional[int] = 0,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> "DeploymentResourcePool":
         """Creates a new DeploymentResourcePool.
 
@@ -209,6 +211,13 @@ def create(
             multihost_gpu_node_count (int):
                 Optional. The number of nodes per replica for multihost GPU
                 deployments. Required for multihost GPU deployments.
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use FlexStart VM powered by
+                Dynamic Workload Scheduler to schedule the deployment workload.
+                The FlexStart VM will be available for a maximum of 7 days or up
+                to the max_runtime_duration specified, whichever is shorter.
+                After this period, the model will be automatically undeployed.
+                The value is in seconds.
 
         Returns:
             DeploymentResourcePool
@@ -237,6 +246,7 @@ def create(
             create_request_timeout=create_request_timeout,
             required_replica_count=required_replica_count,
             multihost_gpu_node_count=multihost_gpu_node_count,
+            max_runtime_duration=max_runtime_duration,
         )
 
     @classmethod
@@ -260,6 +270,7 @@ def _create(
         create_request_timeout: Optional[float] = None,
         required_replica_count: Optional[int] = 0,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> "DeploymentResourcePool":
         """Creates a new DeploymentResourcePool.
 
@@ -332,7 +343,13 @@ def _create(
             multihost_gpu_node_count (int):
                 Optional. The number of nodes per replica for multihost GPU
                 deployments. Required for multihost GPU deployments.
-
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use FlexStart VM powered by Dynamic
+                Workload Scheduler to schedule the deployment workload. The DWS
+                resource will be available for a maximum of 7 days or up to the
+                max_runtime_duration specified, whichever is shorter. After this
+                period, the model will be automatically undeployed. The value is
+                in seconds.
         Returns:
             DeploymentResourcePool
         """
@@ -347,6 +364,10 @@ def _create(
             required_replica_count=required_replica_count,
         )
 
+        prediction_utils.add_flex_start_to_dedicated_resources(
+            dedicated_resources, max_runtime_duration
+        )
+
         machine_spec = gca_machine_resources_compat.MachineSpec(
             machine_type=machine_type,
             multihost_gpu_node_count=multihost_gpu_node_count,
@@ -706,6 +727,7 @@ def deploy(
         required_replica_count: Optional[int] = 0,
         rollout_options: Optional[RolloutOptions] = None,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> None:
         """Deploys a Model to the Endpoint.
 
@@ -809,6 +831,12 @@ def deploy(
             multihost_gpu_node_count (int): Optional. The number of nodes per
               replica for multihost GPU deployments. Required for multihost GPU
               deployments.
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use DWS resource
+                to schedule the deployment workload. The DWS resource will be available
+                for a maximum of 7 days or up to the max_runtime_duration specified,
+                whichever is shorter. After this period, the model will be
+                automatically undeployed. The value is in seconds.
         """
         self._sync_gca_resource_if_skipped()
 
@@ -853,6 +881,7 @@ def deploy(
             required_replica_count=required_replica_count,
             rollout_options=rollout_options,
             multihost_gpu_node_count=multihost_gpu_node_count,
+            max_runtime_duration=max_runtime_duration,
         )
 
     @base.optional_sync()
@@ -882,6 +911,7 @@ def _deploy(
         required_replica_count: Optional[int] = 0,
         rollout_options: Optional[RolloutOptions] = None,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> None:
         """Deploys a Model to the Endpoint.
 
@@ -979,7 +1009,12 @@ def _deploy(
             multihost_gpu_node_count (int): Optional. The number of nodes per
               replica for multihost GPU deployments. Required for multihost
               GPU deployments.
-
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use DWS resource
+                to schedule the deployment workload. The DWS resource will be available
+                for a maximum of 7 days or up to the max_runtime_duration specified,
+                whichever is shorter. After this period, the model will be
+                automatically undeployed. The value is in seconds.
         """
         _LOGGER.log_action_start_against_resource(
             f"Deploying Model {model.resource_name} to", "", self
@@ -1013,6 +1048,7 @@ def _deploy(
             required_replica_count=required_replica_count,
             rollout_options=rollout_options,
             multihost_gpu_node_count=multihost_gpu_node_count,
+            max_runtime_duration=max_runtime_duration,
         )
 
         _LOGGER.log_action_completed_against_resource("model", "deployed", self)
@@ -1049,6 +1085,7 @@ def _deploy_call(
         required_replica_count: Optional[int] = 0,
         rollout_options: Optional[RolloutOptions] = None,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> None:
         """Helper method to deploy model to endpoint.
 
@@ -1153,6 +1190,12 @@ def _deploy_call(
             multihost_gpu_node_count (int):
               Optional. The number of nodes per replica for multihost GPU
               deployments. Required for multihost GPU deployments.
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use DWS resource
+                to schedule the deployment workload. The DWS resource will be available
+                for a maximum of 7 days or up to the max_runtime_duration specified,
+                whichever is shorter. After this period, the model will be
+                automatically undeployed. The value is in seconds.
 
         Raises:
             ValueError: If only `accelerator_type` or `accelerator_count` is
@@ -1235,6 +1278,10 @@ def _deploy_call(
                     required_replica_count=required_replica_count,
                 )
 
+                prediction_utils.add_flex_start_to_dedicated_resources(
+                    dedicated_resources, max_runtime_duration
+                )
+
                 machine_spec = gca_machine_resources_compat.MachineSpec(
                     machine_type=machine_type,
                     multihost_gpu_node_count=multihost_gpu_node_count,
@@ -1599,6 +1646,7 @@ def deploy(
         required_replica_count: Optional[int] = 0,
         rollout_options: Optional[RolloutOptions] = None,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> Union[Endpoint, models.PrivateEndpoint]:
         """Deploys model to endpoint.
 
@@ -1723,7 +1771,12 @@ def deploy(
             multihost_gpu_node_count (int):
               Optional. The number of nodes per replica for multihost GPU
               deployments. Required for multihost GPU deployments.
-
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use DWS resource
+                to schedule the deployment workload. The DWS resource will be available
+                for a maximum of 7 days or up to the max_runtime_duration specified,
+                whichever is shorter. After this period, the model will be
+                automatically undeployed. The value is in seconds.
         Returns:
             endpoint (Union[Endpoint, models.PrivateEndpoint]):
                 Endpoint with the deployed model.
@@ -1785,6 +1838,7 @@ def deploy(
             required_replica_count=required_replica_count,
             rollout_options=rollout_options,
             multihost_gpu_node_count=multihost_gpu_node_count,
+            max_runtime_duration=max_runtime_duration,
         )
 
     def _should_enable_dedicated_endpoint(self, fast_tryout_enabled: bool) -> bool:
@@ -1823,6 +1877,7 @@ def _deploy(
         required_replica_count: Optional[int] = 0,
         rollout_options: Optional[RolloutOptions] = None,
         multihost_gpu_node_count: Optional[int] = None,
+        max_runtime_duration: Optional[int] = None,
     ) -> Union[Endpoint, models.PrivateEndpoint]:
         """Deploys model to endpoint.
 
@@ -1938,6 +1993,12 @@ def _deploy(
             multihost_gpu_node_count (int):
               Optional. The number of nodes per replica for multihost GPU
               deployments. Required for multihost GPU deployments.
+            max_runtime_duration (int):
+                Optional. Immutable. If set, use DWS resource
+                to schedule the deployment workload. The DWS resource will be available
+                for a maximum of 7 days or up to the max_runtime_duration specified,
+                whichever is shorter. After this period, the model will be
+                automatically undeployed. The value is in seconds.
         Returns:
             endpoint (Union[Endpoint, models.PrivateEndpoint]):
                 Endpoint with the deployed model.
@@ -2004,6 +2065,7 @@ def _deploy(
             system_labels=system_labels,
             required_replica_count=required_replica_count,
             multihost_gpu_node_count=multihost_gpu_node_count,
+            max_runtime_duration=max_runtime_duration,
             **preview_kwargs,
         )
 

google/cloud/aiplatform/utils/prediction_utils.py

@@ -26,6 +26,10 @@
 from google.cloud import storage
 from google.cloud.aiplatform.constants import prediction
 from google.cloud.aiplatform.utils import path_utils
+from google.cloud.aiplatform.compat.types import (
+    machine_resources_v1beta1 as gca_machine_resources_compat,
+)
+from google.protobuf import duration_pb2
 
 _logger = logging.getLogger(__name__)
 
@@ -151,3 +155,22 @@ def download_model_artifacts(artifact_uri: str) -> None:
     else:
         # Copy files to the current working directory.
         shutil.copytree(artifact_uri, ".", dirs_exist_ok=True)
+
+
+def add_flex_start_to_dedicated_resources(
+    dedicated_resources: gca_machine_resources_compat.DedicatedResources,
+    max_runtime_duration: Optional[int] = None,
+) -> None:
+    """Adds FlexStart configuration to DedicatedResources if max_runtime_duration is provided.
+
+    Args:
+        dedicated_resources (gca_machine_resources_compat.DedicatedResources):
+            Required. The DedicatedResources object to modify.
+        max_runtime_duration (int):
+            Optional. The maximum runtime duration in seconds. If provided and
+            greater than 0, a FlexStart configuration will be added.
+    """
+    if max_runtime_duration is not None and max_runtime_duration > 0:
+        dedicated_resources.flex_start = gca_machine_resources_compat.FlexStart(
+            max_runtime_duration=duration_pb2.Duration(seconds=max_runtime_duration)
+        )

tests/unit/aiplatform/test_endpoints.py

@@ -279,6 +279,8 @@
     inference_timeout=duration_pb2.Duration(seconds=_TEST_INFERENCE_TIMEOUT)
 )
 
+_TEST_MAX_RUNTIME_DURATION = 600
+
 """
 ----------------------------------------------------------------------------
 Endpoint Fixtures
@@ -4645,3 +4647,60 @@ def test_construct_sdk_resource_from_gapic_uses_resource_project(self):
         )
         assert endpoint2.project != _TEST_PROJECT
         assert endpoint2.location != _TEST_LOCATION
+
+    @pytest.mark.usefixtures(
+        "get_endpoint_mock", "get_model_mock", "create_endpoint_mock"
+    )
+    @pytest.mark.parametrize("sync", [True, False])
+    def test_deploy_no_endpoint_with_max_runtime_duration(
+        self, preview_deploy_model_mock, sync
+    ):
+        test_model = preview_models.Model(_TEST_ID)
+        test_model._gca_resource.supported_deployment_resources_types.append(
+            aiplatform.gapic.Model.DeploymentResourcesType.DEDICATED_RESOURCES
+        )
+        test_endpoint = test_model.deploy(
+            machine_type=_TEST_MACHINE_TYPE,
+            accelerator_type=_TEST_ACCELERATOR_TYPE,
+            accelerator_count=_TEST_ACCELERATOR_COUNT,
+            service_account=_TEST_SERVICE_ACCOUNT,
+            sync=sync,
+            deploy_request_timeout=None,
+            max_runtime_duration=_TEST_MAX_RUNTIME_DURATION,
+        )
+
+        if not sync:
+            test_endpoint.wait()
+
+        expected_machine_spec = gca_machine_resources_v1beta1.MachineSpec(
+            machine_type=_TEST_MACHINE_TYPE,
+            accelerator_type=_TEST_ACCELERATOR_TYPE,
+            accelerator_count=_TEST_ACCELERATOR_COUNT,
+        )
+        expected_dedicated_resources = gca_machine_resources_v1beta1.DedicatedResources(
+            machine_spec=expected_machine_spec,
+            min_replica_count=1,
+            max_replica_count=1,
+            spot=False,
+            required_replica_count=0,
+            flex_start=gca_machine_resources_v1beta1.FlexStart(
+                max_runtime_duration=duration_pb2.Duration(
+                    seconds=_TEST_MAX_RUNTIME_DURATION
+                ),
+            ),
+        )
+        expected_deployed_model = gca_endpoint_v1beta1.DeployedModel(
+            dedicated_resources=expected_dedicated_resources,
+            model=test_model.resource_name,
+            display_name=None,
+            service_account=_TEST_SERVICE_ACCOUNT,
+            enable_container_logging=True,
+            faster_deployment_config=gca_endpoint_v1beta1.FasterDeploymentConfig(),
+        )
+        preview_deploy_model_mock.assert_called_once_with(
+            endpoint=test_endpoint.resource_name,
+            deployed_model=expected_deployed_model,
+            traffic_split={"0": 100},
+            metadata=(),
+            timeout=None,
+        )

tests/unit/aiplatform/test_models.py

@@ -2556,6 +2556,63 @@ def test_preview_deploy_no_endpoint_dedicated_resources_autoscaling_request_coun
             timeout=None,
         )
 
+    @pytest.mark.usefixtures(
+        "get_endpoint_mock",
+        "get_model_mock",
+        "create_endpoint_mock",
+        "preview_deploy_model_mock",
+    )
+    @pytest.mark.parametrize("sync", [True, False])
+    def test_preview_deploy_no_endpoint_with_flex_start(
+        self, preview_deploy_model_mock, sync
+    ):
+        test_model = preview_models.Model(_TEST_ID)
+        test_model._gca_resource.supported_deployment_resources_types.append(
+            aiplatform.gapic.Model.DeploymentResourcesType.DEDICATED_RESOURCES
+        )
+
+        test_endpoint = test_model.deploy(
+            machine_type=_TEST_MACHINE_TYPE,
+            accelerator_type=_TEST_ACCELERATOR_TYPE,
+            accelerator_count=_TEST_ACCELERATOR_COUNT,
+            sync=sync,
+            deploy_request_timeout=None,
+            max_runtime_duration=600,
+        )
+
+        if not sync:
+            test_endpoint.wait()
+
+        expected_machine_spec = gca_machine_resources_v1beta1.MachineSpec(
+            machine_type=_TEST_MACHINE_TYPE,
+            accelerator_type=_TEST_ACCELERATOR_TYPE,
+            accelerator_count=_TEST_ACCELERATOR_COUNT,
+        )
+        expected_dedicated_resources = gca_machine_resources_v1beta1.DedicatedResources(
+            machine_spec=expected_machine_spec,
+            min_replica_count=1,
+            max_replica_count=1,
+            spot=False,
+            flex_start=gca_machine_resources_v1beta1.FlexStart(
+                max_runtime_duration=duration_pb2.Duration(seconds=600),
+            ),
+        )
+        expected_deployed_model = gca_endpoint_v1beta1.DeployedModel(
+            dedicated_resources=expected_dedicated_resources,
+            model=test_model.resource_name,
+            display_name=None,
+            enable_container_logging=True,
+            faster_deployment_config=gca_endpoint_v1beta1.FasterDeploymentConfig(),
+        )
+
+        preview_deploy_model_mock.assert_called_once_with(
+            endpoint=test_endpoint.resource_name,
+            deployed_model=expected_deployed_model,
+            traffic_split={"0": 100},
+            metadata=(),
+            timeout=None,
+        )
+
     @pytest.mark.usefixtures(
         "get_endpoint_mock", "get_model_mock", "create_endpoint_mock"
     )