def __init__(self, scope: core.Construct, id: str, **kwargs) -> None:
super().__init__(scope, id, **kwargs)
model_artefacts_path = './bert_model_artefacts'
bucket_name = 'sagemaker-endpoint-artefacts'
file_name = 'bert_model_artefacts/model.tar.gz'
aws_region = 'eu-west-1'
create_model_artefacts(model_artefacts_path, bucket_name, file_name,
aws_region)
sagemaker_role = aws_iam.Role(
self,
id='SageMaker-BERT-Endpoint-Role',
assumed_by=aws_iam.ServicePrincipal('sagemaker.amazonaws.com'))
sagemaker_role.add_managed_policy(
aws_iam.ManagedPolicy.from_aws_managed_policy_name(
'AmazonSageMakerFullAccess'))
sagemaker_model = aws_sagemaker.CfnModel(
self,
id='SageMaker-BERT-Model',
execution_role_arn=sagemaker_role.role_arn,
primary_container=aws_sagemaker.CfnModel.
ContainerDefinitionProperty(
image=
'763104351884.dkr.ecr.eu-west-1.amazonaws.com/pytorch-inference:1.4.0-cpu-py36-ubuntu16.04',
model_data_url='s3://' + bucket_name + '/' + file_name,
environment={'SAGEMAKER_PROGRAM': 'predictor.py'}),
model_name='bert-model')
sagemaker_endpoint_config = aws_sagemaker.CfnEndpointConfig(
self,
id='SageMaker-BERT-Endpoint-Config',
production_variants=[
aws_sagemaker.CfnEndpointConfig.ProductionVariantProperty(
initial_instance_count=1,
initial_variant_weight=1.0,
instance_type='ml.t2.xlarge',
model_name=sagemaker_model.model_name,
variant_name='bert-model')
],
endpoint_config_name='bert-model-endpoint-config')
sagemaker_endpoint_config.add_depends_on(sagemaker_model)
sagemaker_model_endpoint = aws_sagemaker.CfnEndpoint(
self,
id='SageMaker-BERT-Endpoint',
endpoint_config_name='bert-model-endpoint-config',
endpoint_name='bert-model-endpoint')
sagemaker_model_endpoint.add_depends_on(sagemaker_endpoint_config)
def create_sagemaker_endpoint(scope, id, endpoint_config_name, model_name,
**kwargs):
# create Sagemaker endpoint
sagemaker_endpoint = sagemaker.CfnEndpoint(
scope,
id,
endpoint_config_name=endpoint_config_name,
tags=[{
"key": "endpoint-name",
"value": f"{model_name}-endpoint"
}],
**kwargs,
)
return sagemaker_endpoint
def __init__(self, app: core.App, id: str, **kwargs) -> None:
timestamp = '-' + time.strftime("%Y-%m-%d-%H-%M-%S", time.gmtime())
super().__init__(app, id, **kwargs)
model = sagemaker.CfnModel(
scope=self,
id="my_model",
execution_role_arn=self.node.try_get_context("role_arn"),
containers=[{
"image":
self.node.try_get_context("image"),
"modelDataUrl":
self.node.try_get_context("model_data_url")
}],
model_name=self.node.try_get_context("model_name") + timestamp)
epc = sagemaker.CfnEndpointConfig(
scope=self,
id="my_epc",
production_variants=[{
"modelName":
model.model_name,
"variantName":
"variant-1",
"initialVariantWeight":
1.0,
"initialInstanceCount":
1,
"instanceType":
self.node.try_get_context("instance_type")
}],
endpoint_config_name=self.node.try_get_context("epc_name") +
timestamp)
epc.add_depends_on(model)
ep = sagemaker.CfnEndpoint(
scope=self,
id="my_ep",
endpoint_config_name=epc.endpoint_config_name,
endpoint_name=self.node.try_get_context("ep_name") + timestamp)
ep.add_depends_on(epc)
def __init__(
self,
scope: core.Construct,
construct_id: str,
deployment_config: DeploymentConfig,
project_name: str,
endpoint_name: str,
tags: list,
**kwargs,
) -> None:
super().__init__(scope, construct_id, **kwargs)
# Define the package group names for champion and challenger
champion_package_group = f"{project_name}-champion"
challenger_package_group = f"{project_name}-challenger"
# Get the approved packages for the project
registry = ModelRegistry()
challenger_creation_time: datetime = None
# If we don't have a specific champion variant defined, get the latest approved
if deployment_config.champion_variant_config is None:
logger.info("Selecting top champion variant")
p = registry.get_latest_approved_packages(champion_package_group,
max_results=1)[0]
deployment_config.champion_variant_config = VariantConfig(
model_package_version=p["ModelPackageVersion"],
model_package_arn=p["ModelPackageArn"],
initial_variant_weight=1,
instance_count=deployment_config.instance_count,
instance_type=deployment_config.instance_type,
)
challenger_creation_time = p["CreationTime"]
else:
# Get the versioned package and update ARN
version = deployment_config.champion_variant_config.model_package_version
logger.info(f"Selecting champion version {version}")
p = registry.get_versioned_approved_packages(
champion_package_group,
model_package_versions=[version],
)[0]
deployment_config.champion_variant_config.model_package_arn = p[
"ModelPackageArn"]
# If we don't have challenger variant config, get the latest after challenger creation time
if deployment_config.challenger_variant_config is None:
logger.info(
f"Selecting top {deployment_config.challenger_variant_count} challenger variants created after {challenger_creation_time}"
)
deployment_config.challenger_variant_config = [
VariantConfig(
model_package_version=p["ModelPackageVersion"],
model_package_arn=p["ModelPackageArn"],
initial_variant_weight=1,
instance_count=deployment_config.instance_count,
instance_type=deployment_config.instance_type,
) for p in registry.get_latest_approved_packages(
challenger_package_group,
max_results=deployment_config.challenger_variant_count,
creation_time_after=challenger_creation_time,
)
]
else:
# Get the versioned packages and update ARN
versions = [
c.model_package_version
for c in deployment_config.challenger_variant_config
]
logger.info(f"Selecting challenger versions {versions}")
ps = registry.get_versioned_approved_packages(
challenger_package_group,
model_package_versions=versions,
)
for i, vc in enumerate(
deployment_config.challenger_variant_config):
vc.model_package_arn = ps[i]["ModelPackageArn"]
# Get the service catalog role
service_catalog_role = aws_iam.Role.from_role_arn(
self,
"SageMakerRole",
f"arn:aws:iam::{self.account}:role/service-role/AmazonSageMakerServiceCatalogProductsUseRole",
)
# Add the challenger variant
model_configs = [deployment_config.champion_variant_config
] + deployment_config.challenger_variant_config
model_variants = []
for i, variant_config in enumerate(model_configs):
# If variant name not in config use "Champion" for the latest approved and "Challenge{N}" for next N pending
variant_name = variant_config.variant_name or (
f"Champion{variant_config.model_package_version}" if i == 0
else f"Challenger{variant_config.model_package_version}")
# Do not use a custom named resource for models as these get replaced
model = aws_sagemaker.CfnModel(
self,
variant_name,
execution_role_arn=service_catalog_role.role_arn,
primary_container=aws_sagemaker.CfnModel.
ContainerDefinitionProperty(
model_package_name=variant_config.model_package_arn, ),
)
# Create the production variant
model_variant = aws_sagemaker.CfnEndpointConfig.ProductionVariantProperty(
initial_instance_count=variant_config.instance_count,
initial_variant_weight=variant_config.initial_variant_weight,
instance_type=variant_config.instance_type,
model_name=model.attr_model_name,
variant_name=variant_name,
)
model_variants.append(model_variant)
if len(model_variants) == 0:
raise Exception("No model variants matching configuration")
endpoint_config = aws_sagemaker.CfnEndpointConfig(
self,
"EndpointConfig",
production_variants=model_variants,
)
self.endpoint = aws_sagemaker.CfnEndpoint(
self,
"Endpoint",
endpoint_config_name=endpoint_config.attr_endpoint_config_name,
endpoint_name=endpoint_name,
tags=tags,
)