Skip to content

Commit

Permalink
feat: Adding fast_tryout_enabled option to Vertex public preview SDK
Browse files Browse the repository at this point in the history 
PiperOrigin-RevId: 693029103
  • Loading branch information
@vertex-sdk-bot @copybara-github
vertex-sdk-bot authored and copybara-github committed Nov 7, 2024
1 parent 16f7beb commit c851b52
Show file tree
Hide file tree
Showing 3 changed files with 263 additions and 55 deletions.
215 changes: 160 additions & 55 deletions google/cloud/aiplatform/preview/models.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -518,13 +518,14 @@ def _validate_deploy_args(
"Minimum and maximum replica counts must not be "
"if not using a shared resource pool."
)
return super()._validate_deploy_args(
return aiplatform.Endpoint._validate_deploy_args(
min_replica_count=min_replica_count,
max_replica_count=max_replica_count,
accelerator_type=accelerator_type,
deployed_model_display_name=deployed_model_display_name,
traffic_split=traffic_split,
traffic_percentage=traffic_percentage,
deployment_resource_pool=deployment_resource_pool,
)

if (
Expand Down Expand Up @@ -580,6 +581,7 @@ def deploy(
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
deployment_resource_pool: Optional[DeploymentResourcePool] = None,
disable_container_logging: bool = False,
fast_tryout_enabled: bool = False,
) -> None:
"""Deploys a Model to the Endpoint.
Expand Down Expand Up @@ -661,6 +663,10 @@ def deploy(
disable_container_logging (bool):
If True, container logs from the deployed model will not be
written to Cloud Logging. Defaults to False.
fast_tryout_enabled (bool): Optional.
If True, model will be deployed using faster deployment path.
Useful for quick experiments. Not for production workloads. Only
available for most popular models and machine types. Defaults to False.
"""
self._sync_gca_resource_if_skipped()
Expand Down Expand Up @@ -699,6 +705,7 @@ def deploy(
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
deployment_resource_pool=deployment_resource_pool,
disable_container_logging=disable_container_logging,
fast_tryout_enabled=fast_tryout_enabled,
)

@base.optional_sync()
Expand All @@ -722,6 +729,7 @@ def _deploy(
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
deployment_resource_pool: Optional[DeploymentResourcePool] = None,
disable_container_logging: bool = False,
fast_tryout_enabled: bool = False,
) -> None:
"""Deploys a Model to the Endpoint.
Expand Down Expand Up @@ -797,6 +805,10 @@ def _deploy(
disable_container_logging (bool):
If True, container logs from the deployed model will not be
written to Cloud Logging. Defaults to False.
fast_tryout_enabled (bool): Optional.
If True, model will be deployed using faster deployment path.
Useful for quick experiments. Not for production workloads. Only
available for most popular models and machine types. Defaults to False.
"""
_LOGGER.log_action_start_against_resource(
Expand Down Expand Up @@ -825,6 +837,7 @@ def _deploy(
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
deployment_resource_pool=deployment_resource_pool,
disable_container_logging=disable_container_logging,
fast_tryout_enabled=fast_tryout_enabled,
)

_LOGGER.log_action_completed_against_resource("model", "deployed", self)
Expand Down Expand Up @@ -855,6 +868,7 @@ def _deploy_call(
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
deployment_resource_pool: Optional[DeploymentResourcePool] = None,
disable_container_logging: bool = False,
fast_tryout_enabled: bool = False,
) -> None:
"""Helper method to deploy model to endpoint.
Expand Down Expand Up @@ -937,6 +951,10 @@ def _deploy_call(
disable_container_logging (bool):
If True, container logs from the deployed model will not be
written to Cloud Logging. Defaults to False.
fast_tryout_enabled (bool): Optional.
If True, model will be deployed using faster deployment path.
Useful for quick experiments. Not for production workloads. Only
available for most popular models and machine types. Defaults to False.
Raises:
ValueError: If only `accelerator_type` or `accelerator_count` is
Expand All @@ -950,72 +968,148 @@ def _deploy_call(
are present.
"""
if not deployment_resource_pool:
return super()._deploy_call(
api_client=api_client,
endpoint_resource_name=endpoint_resource_name,
model=model,
endpoint_resource_traffic_split=endpoint_resource_traffic_split,
network=network,
deployed_model_display_name=deployed_model_display_name,
traffic_percentage=traffic_percentage,
traffic_split=traffic_split,
machine_type=machine_type,
min_replica_count=min_replica_count,
max_replica_count=max_replica_count,
accelerator_type=accelerator_type,
accelerator_count=accelerator_count,
max_replica_count = max(min_replica_count, max_replica_count)

if bool(accelerator_type) != bool(accelerator_count):
raise ValueError(
"Both `accelerator_type` and `accelerator_count` should be specified or None."
)

if autoscaling_target_accelerator_duty_cycle is not None and (
not accelerator_type or not accelerator_count
):
raise ValueError(
"Both `accelerator_type` and `accelerator_count` should be set "
"when specifying autoscaling_target_accelerator_duty_cycle`"
)

deployed_model = gca_endpoint_compat.DeployedModel(
model=model.versioned_resource_name,
display_name=deployed_model_display_name,
service_account=service_account,
explanation_spec=explanation_spec,
metadata=metadata,
deploy_request_timeout=deploy_request_timeout,
autoscaling_target_cpu_utilization=autoscaling_target_cpu_utilization,
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
disable_container_logging=disable_container_logging,
enable_container_logging=not disable_container_logging,
)

deployed_model = gca_endpoint_compat.DeployedModel(
model=model.versioned_resource_name,
display_name=deployed_model_display_name,
service_account=service_account,
enable_container_logging=not disable_container_logging,
)

supports_shared_resources = (
gca_model_compat.Model.DeploymentResourcesType.SHARED_RESOURCES
in model.supported_deployment_resources_types
)
supports_automatic_resources = (
gca_model_compat.Model.DeploymentResourcesType.AUTOMATIC_RESOURCES
in model.supported_deployment_resources_types
)
supports_dedicated_resources = (
gca_model_compat.Model.DeploymentResourcesType.DEDICATED_RESOURCES
in model.supported_deployment_resources_types
)
provided_custom_machine_spec = (
machine_type
or accelerator_type
or accelerator_count
or autoscaling_target_accelerator_duty_cycle
or autoscaling_target_cpu_utilization
)

if not supports_shared_resources:
raise ValueError(
"`deployment_resource_pool` may only be specified for models "
" which support shared resources."
# If the model supports both automatic and dedicated deployment resources,
# decide based on the presence of machine spec customizations
use_dedicated_resources = supports_dedicated_resources and (
not supports_automatic_resources or provided_custom_machine_spec
)

provided_custom_machine_spec = (
machine_type
or accelerator_type
or accelerator_count
or autoscaling_target_accelerator_duty_cycle
or autoscaling_target_cpu_utilization
)
if provided_custom_machine_spec and not use_dedicated_resources:
_LOGGER.info(
"Model does not support dedicated deployment resources. "
"The machine_type, accelerator_type and accelerator_count,"
"autoscaling_target_accelerator_duty_cycle,"
"autoscaling_target_cpu_utilization parameters are ignored."
)

if provided_custom_machine_spec:
raise ValueError(
"Conflicting parameters in deployment request. "
"The machine_type, accelerator_type and accelerator_count,"
"autoscaling_target_accelerator_duty_cycle,"
"autoscaling_target_cpu_utilization parameters may not be set "
"when `deployment_resource_pool` is specified."
if use_dedicated_resources and not machine_type:
machine_type = _DEFAULT_MACHINE_TYPE
_LOGGER.info(f"Using default machine_type: {machine_type}")

if use_dedicated_resources:
dedicated_resources = gca_machine_resources_compat.DedicatedResources(
min_replica_count=min_replica_count,
max_replica_count=max_replica_count,
)

machine_spec = gca_machine_resources_compat.MachineSpec(
machine_type=machine_type
)

if autoscaling_target_cpu_utilization:
autoscaling_metric_spec = gca_machine_resources_compat.AutoscalingMetricSpec(
metric_name="aiplatform.googleapis.com/prediction/online/cpu/utilization",
target=autoscaling_target_cpu_utilization,
)
dedicated_resources.autoscaling_metric_specs.extend(
[autoscaling_metric_spec]
)

if accelerator_type and accelerator_count:
utils.validate_accelerator_type(accelerator_type)
machine_spec.accelerator_type = accelerator_type
machine_spec.accelerator_count = accelerator_count

if autoscaling_target_accelerator_duty_cycle:
autoscaling_metric_spec = gca_machine_resources_compat.AutoscalingMetricSpec(
metric_name="aiplatform.googleapis.com/prediction/online/accelerator/duty_cycle",
target=autoscaling_target_accelerator_duty_cycle,
)
dedicated_resources.autoscaling_metric_specs.extend(
[autoscaling_metric_spec]
)

dedicated_resources.machine_spec = machine_spec

# Checking if flag fast_tryout_enabled is set, only in v1beta1
deployed_model.faster_deployment_config = (
gca_endpoint_compat.FasterDeploymentConfig(
fast_tryout_enabled=fast_tryout_enabled
)
)
deployed_model.dedicated_resources = dedicated_resources
else:
deployed_model = gca_endpoint_compat.DeployedModel(
model=model.versioned_resource_name,
display_name=deployed_model_display_name,
service_account=service_account,
enable_container_logging=not disable_container_logging,
)

deployed_model.shared_resources = deployment_resource_pool.resource_name
supports_shared_resources = (
gca_model_compat.Model.DeploymentResourcesType.SHARED_RESOURCES
in model.supported_deployment_resources_types
)

if explanation_spec:
raise ValueError(
"Model explanation is not supported for deployments using "
"shared resources."
if not supports_shared_resources:
raise ValueError(
"`deployment_resource_pool` may only be specified for models "
" which support shared resources."
)

provided_custom_machine_spec = (
machine_type
or accelerator_type
or accelerator_count
or autoscaling_target_accelerator_duty_cycle
or autoscaling_target_cpu_utilization
)

if provided_custom_machine_spec:
raise ValueError(
"Conflicting parameters in deployment request. "
"The machine_type, accelerator_type and accelerator_count,"
"autoscaling_target_accelerator_duty_cycle,"
"autoscaling_target_cpu_utilization parameters may not be set "
"when `deployment_resource_pool` is specified."
)

deployed_model.shared_resources = deployment_resource_pool.resource_name

if explanation_spec:
raise ValueError(
"Model explanation is not supported for deployments using "
"shared resources."
)

# Checking if traffic percentage is valid
# TODO(b/221059294) PrivateEndpoint should support traffic split
if traffic_split is None and not network:
Expand Down Expand Up @@ -1264,6 +1358,7 @@ def deploy(
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
deployment_resource_pool: Optional[DeploymentResourcePool] = None,
disable_container_logging: bool = False,
fast_tryout_enabled: bool = False,
) -> Union[Endpoint, models.PrivateEndpoint]:
"""Deploys model to endpoint.
Expand Down Expand Up @@ -1366,6 +1461,10 @@ def deploy(
disable_container_logging (bool):
If True, container logs from the deployed model will not be
written to Cloud Logging. Defaults to False.
fast_tryout_enabled (bool): Optional.
If True, model will be deployed using faster deployment path.
Useful for quick experiments. Not for production workloads. Only
available for most popular models and machine types. Defaults to False.
Returns:
endpoint (Union[Endpoint, models.PrivateEndpoint]):
Expand Down Expand Up @@ -1421,6 +1520,7 @@ def deploy(
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
deployment_resource_pool=deployment_resource_pool,
disable_container_logging=disable_container_logging,
fast_tryout_enabled=fast_tryout_enabled,
)

@base.optional_sync(return_input_arg="endpoint", bind_future_to_self=False)
Expand All @@ -1446,6 +1546,7 @@ def _deploy(
autoscaling_target_accelerator_duty_cycle: Optional[int] = None,
deployment_resource_pool: Optional[DeploymentResourcePool] = None,
disable_container_logging: bool = False,
fast_tryout_enabled: bool = False,
) -> Union[Endpoint, models.PrivateEndpoint]:
"""Deploys model to endpoint.
Expand Down Expand Up @@ -1540,6 +1641,9 @@ def _deploy(
disable_container_logging (bool):
If True, container logs from the deployed model will not be
written to Cloud Logging. Defaults to False.
fast_tryout_enabled (bool):
Optional. Whether to enable fast deployment. Defaults to False.
Useful for quick experiments. Not for production workloads.
Returns:
endpoint (Union[Endpoint, models.PrivateEndpoint]):
Expand Down Expand Up @@ -1591,6 +1695,7 @@ def _deploy(
autoscaling_target_accelerator_duty_cycle=autoscaling_target_accelerator_duty_cycle,
deployment_resource_pool=deployment_resource_pool,
disable_container_logging=disable_container_logging,
fast_tryout_enabled=fast_tryout_enabled,
)

_LOGGER.log_action_completed_against_resource("model", "deployed", endpoint)
Expand Down
Loading

0 comments on commit c851b52

Please sign in to comment.