CDK python Sagemaker endpoint subnet deletion blocked by ENI - aws-cloudformation

Question

I am deploying with a public and private subnet in two AZs (see code below). However, when I delete I get the following error.
"The subnet 'subnet-0d50f818f269ce4f2' has dependencies and cannot be deleted. (Service: Ec2, Status Code: 400...
When I try to manually delete the subnet it shows that I have a network interfact attached.
Under EC2 > Network instances, I see that the stack has created a nat gateway for each public subnet (2 azs), but an ENI for only one of private subnets (with description SageMaker managed ENI for endpoint mymodelendpoint)? After I delete this ENI the delete process finishes. If I want to delete the ENI manually before cdk destroy has tried and failed I first have to detach it. However, if I run destroy first it is already detached and I can delete manually right away.
Why did it create ENI for only one of the private subnets (I have to azs)? Why do I have to manually delete it? Is there a way to get cdk to delete this as part of the destroy command?
Could it be that cdk destroy somehow doing detach/delete out of order and it fails but leaves the ENI detached?
Thanks
# https://klichx.dev/2021/03/22/sagemaker-deployments-with-aws-cdk/
contrainer_definition_prop = sagemaker.CfnModel.ContainerDefinitionProperty(
image=f'763104351884.dkr.ecr.{REGION}.amazonaws.com/huggingface-pytorch-inference:1.9.0-transformers4.11.0-cpu-py38-ubuntu20.04',
image_config = sagemaker.CfnModel.ImageConfigProperty(
repository_access_mode="Platform"),
container_hostname=f"hf-model-container",
model_data_url=f"s3://sagemaker-{REGION}-{ACCOUNT_ID}/finetune-distilbert-base-cased-2023-02-10-00-55-05/output/model.tar.gz",
)
vpc = ec2.Vpc(
self,
"VPC",
max_azs=2,
subnet_configuration=[
ec2.SubnetConfiguration(
name="public-model-subnet", subnet_type=ec2.SubnetType.PUBLIC
),
ec2.SubnetConfiguration(
name="private-model-subnet", subnet_type=ec2.SubnetType.PRIVATE_WITH_EGRESS
),
],
gateway_endpoints={
"S3": ec2.GatewayVpcEndpointOptions(
service=ec2.GatewayVpcEndpointAwsService.S3
)
},
)
print('vpc.private_subnets', vpc.private_subnets)
model_vpc_config = sagemaker.CfnModel.VpcConfigProperty(
security_group_ids=[vpc.vpc_default_security_group],
subnets=[s.subnet_id for s in vpc.private_subnets],
)
mymodel = sagemaker.CfnModel(
self,
"Inference",
execution_role_arn=self.sm_execution_role.role_arn,
model_name="triage-multi-" + str(int(time.time())),
#containers=[contrainer_definition_prop],
primary_container=contrainer_definition_prop,
# inference_execution_config=sagemaker.CfnModel.InferenceExecutionConfigProperty(
# mode="Direct"
# ),
vpc_config=model_vpc_config
)
model_endpoint_config = sagemaker.CfnEndpointConfig(
self,
"mymodel-endpoint-config",
production_variants=[
sagemaker.CfnEndpointConfig.ProductionVariantProperty(
initial_instance_count=1,
initial_variant_weight=1.0,
instance_type="ml.m5.xlarge",
model_name=mymodel.model_name,
variant_name="production-medium",
),
],
)
model_endpoint_config.add_depends_on(mymodel)
self.model_endpoint = sagemaker.CfnEndpoint(
self,
"mymodel-endpoint", endpoint_config_name=model_endpoint_config.attr_endpoint_config_name,
)