ODSC 68580- Update Evaluation SDK to Support Multi-Model Deployment

ads/aqua/evaluation/evaluation.py

@@ -31,6 +31,7 @@
     Tags,
 )
 from ads.aqua.common.errors import (
+    AquaError,
     AquaFileExistsError,
     AquaFileNotFoundError,
     AquaMissingKeyError,
@@ -75,6 +76,7 @@
     CreateAquaEvaluationDetails,
 )
 from ads.aqua.evaluation.errors import EVALUATION_JOB_EXIT_CODE_MESSAGE
+from ads.aqua.model.constants import ModelCustomMetadataFields
 from ads.aqua.ui import AquaContainerConfig
 from ads.common.auth import default_signer
 from ads.common.object_storage_details import ObjectStorageDetails
@@ -183,6 +185,26 @@ def create(
             evaluation_source = ModelDeployment.from_id(
                 create_aqua_evaluation_details.evaluation_source_id
             )
+            try:
+                if Tags.MULTIMODEL_TYPE_TAG in evaluation_source.freeform_tags:
+                    multi_model_id = evaluation_source.freeform_tags.get(
+                        Tags.AQUA_MODEL_ID_TAG, UNKNOWN
+                    )
+
+                    if not multi_model_id:
+                        raise AquaRuntimeError(
+                            f"Invalid multi model deployment {multi_model_id}."
+                            f"Make sure the {Tags.AQUA_MODEL_ID_TAG} tag is added to the deployment."
+                        )
+
+                    aqua_model = DataScienceModel.from_id(multi_model_id)
+                    AquaEvaluationApp.validate_model_name(
+                        aqua_model, create_aqua_evaluation_details
+                    )
+
+            except AquaError as err:
+                raise AquaValueError(f"{err}") from err
+
             try:
                 if (
                     evaluation_source.runtime.type
@@ -550,6 +572,63 @@ def create(
             parameters=AquaEvalParams(),
         )
 
+    @staticmethod
+    def validate_model_name(
+        evaluation_source: DataScienceModel,
+        create_aqua_evaluation_details: CreateAquaEvaluationDetails,
+    ) -> None:
+        """
+        Validates the user input of the model name when creating an Aqua evaluation.
+
+        Parameters
+        ----------
+        evaluation_source: DataScienceModel
+            The DataScienceModel Object which contains all metadata
+            about each model in a single and multi model deployment.
+        create_aqua_evaluation_details: CreateAquaEvaluationDetails
+            The CreateAquaEvaluationDetails data class which contains all
+            required and optional fields to create the aqua evaluation.
+
+        Raises
+        -------
+        AquaValueError:
+            - When the user fails to specify any input for the model name.
+            - When the user supplies a model name that does not match the model name set in the DataScienceModel metadata."""
+        user_model_parameters = create_aqua_evaluation_details.model_parameters
+
+        custom_metadata_list = evaluation_source.custom_metadata_list
+        user_model_name = user_model_parameters.get("model")
+
+        model_group_count = int(
+            custom_metadata_list.get(
+                ModelCustomMetadataFields.MULTIMODEL_GROUP_COUNT
+            ).value
+        )
+
+        model_names = [
+            custom_metadata_list.get(f"model-name-{idx}").value
+            for idx in range(model_group_count)
+        ]
+
+        valid_model_names = ", ".join(map(str, model_names))
+
+        if "model" not in user_model_parameters:
+            logger.debug(
+                f"User did not input model name for multi model deployment evaluation with evaluation source ID: {create_aqua_evaluation_details.evaluation_source_id}"
+            )
+            raise AquaValueError(
+                f"Provide the model name. For evaluation, a single model needs to be targeted using the name in the multi model deployment. The valid model names for this Model Deployment are {valid_model_names}."
+            )
+
+        if user_model_name not in model_names:
+
+            logger.debug(
+                f"User input for model name was {user_model_name}, expected {valid_model_names} evaluation source ID: {create_aqua_evaluation_details.evaluation_source_id}"
+            )
+            raise AquaValueError(
+                f"Provide the correct model name. The valid model names for this Model Deployment are {valid_model_names}."
+            )
+
     def _build_evaluation_runtime(
         self,
         evaluation_id: str,
@@ -1392,7 +1471,7 @@ def _fetch_jobrun(
             )
         except Exception as e:
             logger.debug(
-                f"Failed to retreive job run: {jobrun_id}. " f"DEBUG INFO: {str(e)}"
+                f"Failed to retreive job run: {jobrun_id}. DEBUG INFO: {str(e)}"
             )
             jobrun = None
 

tests/unitary/with_extras/aqua/test_evaluation.py

@@ -17,6 +17,7 @@
 from ads.aqua.common import utils
 from ads.aqua.common.enums import Tags
 from ads.aqua.common.errors import (
+    AquaError,
     AquaFileNotFoundError,
     AquaMissingKeyError,
     AquaRuntimeError,
@@ -34,6 +35,7 @@
     AquaEvalMetrics,
     AquaEvalReport,
     AquaEvaluationSummary,
+    CreateAquaEvaluationDetails,
 )
 from ads.aqua.extension.base_handler import AquaAPIhandler
 from ads.jobs.ads_job import DataScienceJob, DataScienceJobRun, Job
@@ -353,6 +355,7 @@ class TestDataset:
     COMPARTMENT_ID = "ocid1.compartment.oc1..<UNIQUE_OCID>"
     EVAL_ID = "ocid1.datasciencemodel.oc1.iad.<OCID>"
     INVALID_EVAL_ID = "ocid1.datasciencemodel.oc1.phx.<OCID>"
+    MODEL_DEPLOYMENT_ID = "ocid1.datasciencemodeldeployment.oc1.<region>.<MD_OCID>"
 
 
 class TestAquaEvaluation(unittest.TestCase):
@@ -533,6 +536,67 @@ def test_create_evaluation(
             "time_created": f"{oci_dsc_model.time_created}",
         }
 
+    @parameterized.expand(
+    [
+        (
+            {"model": "model_one"},
+            None
+        ),
+        (
+            {},
+            "Provide the model name. For evaluation, a single model needs to be targeted using the name in the multi model deployment. The valid model names for this Model Deployment are model_one, model_two, model_three."
+        ),
+        (
+            {"model": "wrong_model_name"},
+            "Provide the correct model name. The valid model names for this Model Deployment are model_one, model_two, model_three."
+        )
+    ])
+    @patch("ads.aqua.evaluation.evaluation.AquaEvaluationApp.create")
+    def test_validate_model_name(
+        self,
+        mock_model_parameters,
+        expected_message,
+        mock_model
+    ):
+        curr_dir = os.path.dirname(__file__)
+
+        eval_model_freeform_tags = {"ftag1": "fvalue1", "ftag2": "fvalue2"}
+        eval_model_defined_tags = {"dtag1": "dvalue1", "dtag2": "dvalue2"}
+
+        eval_model_freeform_tags[Tags.MULTIMODEL_TYPE_TAG] = "true"
+        eval_model_freeform_tags[Tags.AQUA_TAG] = "active"
+
+        create_aqua_evaluation_details = dict(  # noqa: C408
+            evaluation_source_id= TestDataset.MODEL_DEPLOYMENT_ID,
+            evaluation_name="test_evaluation_name",
+            dataset_path="oci://dataset_bucket@namespace/prefix/dataset.jsonl",
+            report_path="oci://report_bucket@namespace/prefix/",
+            model_parameters=mock_model_parameters,
+            shape_name="VM.Standard.E3.Flex",
+            block_storage_size=1,
+            experiment_name="test_experiment_name",
+            memory_in_gbs=1,
+            ocpus=1,
+            freeform_tags=eval_model_freeform_tags,
+            defined_tags=eval_model_defined_tags,
+        )
+
+
+        aqua_multi_model = os.path.join(
+            curr_dir, "test_data/deployment/aqua_multi_model.yaml"
+        )
+
+        mock_model = DataScienceModel.from_yaml(
+            uri=aqua_multi_model
+        )
+
+        mock_create_aqua_evaluation_details = MagicMock(**create_aqua_evaluation_details, spec=CreateAquaEvaluationDetails)
+
+        try:
+            AquaEvaluationApp.validate_model_name(mock_model, mock_create_aqua_evaluation_details)
+        except AquaError as e:
+            self.assertEqual(str(e), expected_message)
+
     def test_get_service_model_name(self):
         # get service model name from fine tuned model deployment
         source = ModelDeployment().with_freeform_tags(