feat: Add support for Deployment Recommendation ID in model.deploy(). No tagging support (#3920)

Co-authored-by: jinpengqi <[email protected]>
Co-authored-by: Gary Wang 😤 <[email protected]>

Author: gwang111

src/sagemaker/inference_recommender/inference_recommender_mixin.py

@@ -237,7 +237,12 @@ def _update_params(
                 async_inference_config,
                 explainer_config,
             )
-        return inference_recommendation or (instance_type, initial_instance_count)
+
+        return (
+            inference_recommendation
+            if inference_recommendation
+            else (instance_type, initial_instance_count)
+        )
 
     def _update_params_for_right_size(
         self,
@@ -365,12 +370,6 @@ def _update_params_for_recommendation_id(
             return (instance_type, initial_instance_count)
 
         # Validate non-compatible parameters with recommendation id
-        if bool(instance_type) != bool(initial_instance_count):
-            raise ValueError(
-                "Please either do not specify instance_type and initial_instance_count"
-                "since they are in recommendation, or specify both of them if you want"
-                "to override the recommendation."
-            )
         if accelerator_type is not None:
             raise ValueError("accelerator_type is not compatible with inference_recommendation_id.")
         if async_inference_config is not None:
@@ -386,30 +385,38 @@ def _update_params_for_recommendation_id(
 
         # Validate recommendation id
         if not re.match(r"[a-zA-Z0-9](-*[a-zA-Z0-9]){0,63}\/\w{8}$", inference_recommendation_id):
-            raise ValueError("Inference Recommendation id is not valid")
-        recommendation_job_name = inference_recommendation_id.split("/")[0]
+            raise ValueError("inference_recommendation_id is not valid")
+        job_or_model_name = inference_recommendation_id.split("/")[0]
 
         sage_client = self.sagemaker_session.sagemaker_client
-        recommendation_res = sage_client.describe_inference_recommendations_job(
-            JobName=recommendation_job_name
+        # Get recommendation from right size job and model
+        (
+            right_size_recommendation,
+            model_recommendation,
+            right_size_job_res,
+        ) = self._get_recommendation(
+            sage_client=sage_client,
+            job_or_model_name=job_or_model_name,
+            inference_recommendation_id=inference_recommendation_id,
         )
-        input_config = recommendation_res["InputConfig"]
 
-        recommendation = next(
-            (
-                rec
-                for rec in recommendation_res["InferenceRecommendations"]
-                if rec["RecommendationId"] == inference_recommendation_id
-            ),
-            None,
-        )
+        # Update params beased on model recommendation
+        if model_recommendation:
+            if initial_instance_count is None:
+                raise ValueError("Must specify model recommendation id and instance count.")
+            self.env.update(model_recommendation["Environment"])
+            instance_type = model_recommendation["InstanceType"]
+            return (instance_type, initial_instance_count)
 
-        if not recommendation:
+        # Update params based on default inference recommendation
+        if bool(instance_type) != bool(initial_instance_count):
             raise ValueError(
-                "inference_recommendation_id does not exist in InferenceRecommendations list"
+                "instance_type and initial_instance_count are mutually exclusive with"
+                "recommendation id since they are in recommendation."
+                "Please specify both of them if you want to override the recommendation."
             )
-
-        model_config = recommendation["ModelConfiguration"]
+        input_config = right_size_job_res["InputConfig"]
+        model_config = right_size_recommendation["ModelConfiguration"]
         envs = (
             model_config["EnvironmentParameters"]
             if "EnvironmentParameters" in model_config
@@ -458,8 +465,10 @@ def _update_params_for_recommendation_id(
                 self.model_data = compilation_res["ModelArtifacts"]["S3ModelArtifacts"]
                 self.image_uri = compilation_res["InferenceImage"]
 
-        instance_type = recommendation["EndpointConfiguration"]["InstanceType"]
-        initial_instance_count = recommendation["EndpointConfiguration"]["InitialInstanceCount"]
+        instance_type = right_size_recommendation["EndpointConfiguration"]["InstanceType"]
+        initial_instance_count = right_size_recommendation["EndpointConfiguration"][
+            "InitialInstanceCount"
+        ]
 
         return (instance_type, initial_instance_count)
 
@@ -527,3 +536,77 @@ def _convert_to_stopping_conditions_json(
                 threshold.to_json for threshold in model_latency_thresholds
             ]
         return stopping_conditions
+
+    def _get_recommendation(self, sage_client, job_or_model_name, inference_recommendation_id):
+        """Get recommendation from right size job and model"""
+        right_size_recommendation, model_recommendation, right_size_job_res = None, None, None
+        right_size_recommendation, right_size_job_res = self._get_right_size_recommendation(
+            sage_client=sage_client,
+            job_or_model_name=job_or_model_name,
+            inference_recommendation_id=inference_recommendation_id,
+        )
+        if right_size_recommendation is None:
+            model_recommendation = self._get_model_recommendation(
+                sage_client=sage_client,
+                job_or_model_name=job_or_model_name,
+                inference_recommendation_id=inference_recommendation_id,
+            )
+            if model_recommendation is None:
+                raise ValueError("inference_recommendation_id is not valid")
+
+        return right_size_recommendation, model_recommendation, right_size_job_res
+
+    def _get_right_size_recommendation(
+        self,
+        sage_client,
+        job_or_model_name,
+        inference_recommendation_id,
+    ):
+        """Get recommendation from right size job"""
+        right_size_recommendation, right_size_job_res = None, None
+        try:
+            right_size_job_res = sage_client.describe_inference_recommendations_job(
+                JobName=job_or_model_name
+            )
+            if right_size_job_res:
+                right_size_recommendation = self._search_recommendation(
+                    recommendation_list=right_size_job_res["InferenceRecommendations"],
+                    inference_recommendation_id=inference_recommendation_id,
+                )
+        except sage_client.exceptions.ResourceNotFound:
+            pass
+
+        return right_size_recommendation, right_size_job_res
+
+    def _get_model_recommendation(
+        self,
+        sage_client,
+        job_or_model_name,
+        inference_recommendation_id,
+    ):
+        """Get recommendation from model"""
+        model_recommendation = None
+        try:
+            model_res = sage_client.describe_model(ModelName=job_or_model_name)
+            if model_res:
+                model_recommendation = self._search_recommendation(
+                    recommendation_list=model_res["DeploymentRecommendation"][
+                        "RealTimeInferenceRecommendations"
+                    ],
+                    inference_recommendation_id=inference_recommendation_id,
+                )
+        except sage_client.exceptions.ResourceNotFound:
+            pass
+
+        return model_recommendation
+
+    def _search_recommendation(self, recommendation_list, inference_recommendation_id):
+        """Search recommendation based on recommendation id"""
+        return next(
+            (
+                rec
+                for rec in recommendation_list
+                if rec["RecommendationId"] == inference_recommendation_id
+            ),
+            None,
+        )

src/sagemaker/model.py

@@ -1210,6 +1210,8 @@ def deploy(
             inference_recommendation_id (str): The recommendation id which specifies the
                 recommendation you picked from inference recommendation job results and
                 would like to deploy the model and endpoint with recommended parameters.
+                This can also be a recommendation id returned from ``DescribeModel`` contained in
+                a list of ``RealtimeInferenceRecommendations`` within ``DeploymentRecommendation``
             explainer_config (sagemaker.explainer.ExplainerConfig): Specifies online explainability
                 configuration for use with Amazon SageMaker Clarify. Default: None.
         Raises:

tests/integ/test_inference_recommender.py

@@ -303,6 +303,30 @@ def default_right_sized_unregistered_base_model(sagemaker_session, cpu_instance_
             sagemaker_session.delete_model(ModelName=model.name)
 
 
+@pytest.fixture(scope="module")
+def created_base_model(sagemaker_session, cpu_instance_type):
+    model_data = sagemaker_session.upload_data(path=IR_SKLEARN_MODEL)
+    region = sagemaker_session._region_name
+    image_uri = image_uris.retrieve(
+        framework="sklearn", region=region, version="1.0-1", image_scope="inference"
+    )
+
+    iam_client = sagemaker_session.boto_session.client("iam")
+    role_arn = iam_client.get_role(RoleName="SageMakerRole")["Role"]["Arn"]
+
+    model = Model(
+        model_data=model_data,
+        role=role_arn,
+        entry_point=IR_SKLEARN_ENTRY_POINT,
+        image_uri=image_uri,
+        sagemaker_session=sagemaker_session,
+    )
+
+    model.create(instance_type=cpu_instance_type)
+
+    return model
+
+
 @pytest.mark.slow_test
 def test_default_right_size_and_deploy_registered_model_sklearn(
     default_right_sized_model, sagemaker_session
@@ -453,3 +477,56 @@ def test_deploy_inference_recommendation_id_with_registered_model_sklearn(
             )
             predictor.delete_model()
             predictor.delete_endpoint()
+
+
+@pytest.mark.slow_test
+def test_deploy_deployment_recommendation_id_with_model(created_base_model, sagemaker_session):
+    with timeout(minutes=20):
+        try:
+            deployment_recommendation = poll_for_deployment_recommendation(
+                created_base_model, sagemaker_session
+            )
+
+            assert deployment_recommendation is not None
+
+            real_time_recommendations = deployment_recommendation.get(
+                "RealTimeInferenceRecommendations"
+            )
+            recommendation_id = real_time_recommendations[0].get("RecommendationId")
+
+            endpoint_name = unique_name_from_base("test-rec-id-deployment-default-sklearn")
+            created_base_model.predictor_cls = SKLearnPredictor
+            predictor = created_base_model.deploy(
+                inference_recommendation_id=recommendation_id,
+                initial_instance_count=1,
+                endpoint_name=endpoint_name,
+            )
+
+            payload = pd.read_csv(IR_SKLEARN_DATA, header=None)
+
+            inference = predictor.predict(payload)
+            assert inference is not None
+            assert 26 == len(inference)
+        finally:
+            predictor.delete_model()
+            predictor.delete_endpoint()
+
+
+def poll_for_deployment_recommendation(created_base_model, sagemaker_session):
+    with timeout(minutes=1):
+        try:
+            completed = False
+            while not completed:
+                describe_model_response = sagemaker_session.sagemaker_client.describe_model(
+                    ModelName=created_base_model.name
+                )
+                deployment_recommendation = describe_model_response.get("DeploymentRecommendation")
+
+                completed = (
+                    deployment_recommendation is not None
+                    and "COMPLETED" == deployment_recommendation.get("RecommendationStatus")
+                )
+            return deployment_recommendation
+        except Exception as e:
+            created_base_model.delete_model()
+            raise e

tests/unit/sagemaker/inference_recommender/constants.py

@@ -37,7 +37,10 @@
 
 INVALID_RECOMMENDATION_ID = "ir-job6ab0ff22"
 NOT_EXISTED_RECOMMENDATION_ID = IR_JOB_NAME + "/ad3ec9ee"
+NOT_EXISTED_MODEL_RECOMMENDATION_ID = IR_MODEL_NAME + "/ad3ec9ee"
 RECOMMENDATION_ID = IR_JOB_NAME + "/5bcee92e"
+MODEL_RECOMMENDATION_ID = IR_MODEL_NAME + "/v0KObO5d"
+MODEL_RECOMMENDATION_ENV = {"TS_DEFAULT_WORKERS_PER_MODEL": "4"}
 
 IR_CONTAINER_CONFIG = {
     "Domain": "MACHINE_LEARNING",
@@ -95,6 +98,21 @@
         "Image": IR_IMAGE,
         "ModelDataUrl": IR_MODEL_DATA,
     },
+    "DeploymentRecommendation": {
+        "RecommendationStatus": "COMPLETED",
+        "RealTimeInferenceRecommendations": [
+            {
+                "RecommendationId": MODEL_RECOMMENDATION_ID,
+                "InstanceType": "ml.g4dn.2xlarge",
+                "Environment": MODEL_RECOMMENDATION_ENV,
+            },
+            {
+                "RecommendationId": "test-model-name/d248qVYU",
+                "InstanceType": "ml.c6i.large",
+                "Environment": {},
+            },
+        ],
+    },
 }
 
 DESCRIBE_MODEL_PACKAGE_RESPONSE = {
@@ -134,3 +152,30 @@
     "ModelArtifacts": {"S3ModelArtifacts": IR_COMPILATION_MODEL_DATA},
     "InferenceImage": IR_COMPILATION_IMAGE,
 }
+
+IR_CONTAINER_DEF = {
+    "Image": IR_IMAGE,
+    "Environment": IR_ENV,
+    "ModelDataUrl": IR_MODEL_DATA,
+}
+
+DEPLOYMENT_RECOMMENDATION_CONTAINER_DEF = {
+    "Image": IR_IMAGE,
+    "Environment": MODEL_RECOMMENDATION_ENV,
+    "ModelDataUrl": IR_MODEL_DATA,
+}
+
+IR_COMPILATION_CONTAINER_DEF = {
+    "Image": IR_COMPILATION_IMAGE,
+    "Environment": {},
+    "ModelDataUrl": IR_COMPILATION_MODEL_DATA,
+}
+
+IR_MODEL_PACKAGE_CONTAINER_DEF = {
+    "ModelPackageName": IR_MODEL_PACKAGE_VERSION_ARN,
+    "Environment": IR_ENV,
+}
+
+IR_COMPILATION_MODEL_PACKAGE_CONTAINER_DEF = {
+    "ModelPackageName": IR_MODEL_PACKAGE_VERSION_ARN,
+}