fix: pass name from modelbuilder constructor to created model

src/sagemaker/serve/builder/djl_builder.py

@@ -92,6 +92,7 @@ def __init__(self):
         self.nb_instance_type = None
         self.ram_usage_model_load = None
         self.role_arn = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self):
@@ -130,6 +131,7 @@ def _create_djl_model(self) -> Type[Model]:
             huggingface_hub_token=self.env_vars.get("HF_TOKEN"),
             image_config=self.image_config,
             vpc_config=self.vpc_config,
+            name=self.name,
         )
 
         if not self.image_uri:

src/sagemaker/serve/builder/jumpstart_builder.py

@@ -121,6 +121,7 @@ def __init__(self):
         self.is_compiled = False
         self.is_quantized = False
         self.speculative_decoding_draft_model_source = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self, **kwargs):
@@ -147,7 +148,10 @@ def _is_jumpstart_model_id(self) -> bool:
     def _create_pre_trained_js_model(self) -> Type[Model]:
         """Placeholder docstring"""
         pysdk_model = JumpStartModel(
-            self.model, vpc_config=self.vpc_config, sagemaker_session=self.sagemaker_session
+            self.model,
+            vpc_config=self.vpc_config,
+            sagemaker_session=self.sagemaker_session,
+            name=self.name,
         )
 
         self._original_deploy = pysdk_model.deploy

src/sagemaker/serve/builder/model_builder.py

@@ -492,6 +492,7 @@ def _create_model(self):
             env=self.env_vars,
             sagemaker_session=self.sagemaker_session,
             predictor_cls=self._get_predictor,
+            name=self.name,
         )
 
         # store the modes in the model so that we may

src/sagemaker/serve/builder/tei_builder.py

@@ -65,6 +65,7 @@ def __init__(self):
         self.ram_usage_model_load = None
         self.secret_key = None
         self.role_arn = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self, *args, **kwargs):
@@ -105,6 +106,7 @@ def _create_tei_model(self, **kwargs) -> Type[Model]:
             env=self.env_vars,
             role=self.role_arn,
             sagemaker_session=self.sagemaker_session,
+            name=self.name,
         )
 
         logger.info("Detected %s. Proceeding with the the deployment.", self.image_uri)

src/sagemaker/serve/builder/tf_serving_builder.py

@@ -51,6 +51,7 @@ def __init__(self):
         self.pysdk_model = None
         self.schema_builder = None
         self.env_vars = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self):
@@ -97,6 +98,7 @@ def _create_tensorflow_model(self):
             env=self.env_vars,
             sagemaker_session=self.sagemaker_session,
             predictor_cls=self._get_tensorflow_predictor,
+            name=self.name,
         )
 
         self.pysdk_model.mode = self.mode

src/sagemaker/serve/builder/tgi_builder.py

@@ -92,6 +92,7 @@ def __init__(self):
         self.ram_usage_model_load = None
         self.secret_key = None
         self.role_arn = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self, *args, **kwargs):
@@ -142,6 +143,7 @@ def _create_tgi_model(self) -> Type[Model]:
             env=self.env_vars,
             role=self.role_arn,
             sagemaker_session=self.sagemaker_session,
+            name=self.name,
         )
 
         self._original_deploy = pysdk_model.deploy

src/sagemaker/serve/builder/transformers_builder.py

@@ -89,6 +89,7 @@ def __init__(self):
         self.schema_builder = None
         self.inference_spec = None
         self.shared_libs = None
+        self.name = None
 
     @abstractmethod
     def _prepare_for_mode(self, *args, **kwargs):
@@ -105,6 +106,7 @@ def _create_transformers_model(self) -> Type[Model]:
                 env=self.env_vars,
                 role=self.role_arn,
                 sagemaker_session=self.sagemaker_session,
+                name=self.name,
             )
 
         logger.info("Detected %s. Proceeding with the the deployment.", self.image_uri)

tests/unit/sagemaker/serve/builder/test_djl_builder.py

@@ -78,6 +78,7 @@ def test_build_deploy_for_djl_local_container(
     ):
         builder = ModelBuilder(
             model=mock_model_id,
+            name="mock_model_name",
             schema_builder=mock_schema_builder,
             mode=Mode.LOCAL_CONTAINER,
             model_server=ModelServer.DJL_SERVING,
@@ -89,6 +90,8 @@ def test_build_deploy_for_djl_local_container(
         builder._prepare_for_mode.side_effect = None
 
         model = builder.build()
+        assert model.name == "mock_model_name"
+
         builder.serve_settings.telemetry_opt_out = True
 
         assert isinstance(model, DJLModel)

tests/unit/sagemaker/serve/builder/test_model_builder.py

@@ -317,7 +317,7 @@ def test_build_happy_path_with_sagemaker_endpoint_mode_and_byoc(
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and image_config == MOCK_IMAGE_CONFIG
@@ -326,6 +326,7 @@ def test_build_happy_path_with_sagemaker_endpoint_mode_and_byoc(
             and role == mock_role_arn
             and env == ENV_VARS
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -425,13 +426,14 @@ def test_build_happy_path_with_sagemaker_endpoint_mode_and_1p_dlc_as_byoc(
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_1p_dlc_image_uri
             and model_data == model_data
             and role == mock_role_arn
             and env == ENV_VARS
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -532,13 +534,14 @@ def test_build_happy_path_with_sagemaker_endpoint_mode_and_inference_spec(
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data == model_data
             and role == mock_role_arn
             and env == ENV_VARS_INF_SPEC
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -633,13 +636,14 @@ def test_build_happy_path_with_sagemakerEndpoint_mode_and_model(
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data == model_data
             and role == mock_role_arn
             and env == ENV_VARS
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -742,13 +746,14 @@ def test_build_happy_path_with_sagemakerEndpoint_mode_and_xgboost_model(
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data == model_data
             and role == mock_role_arn
             and env == ENV_VARS
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -847,13 +852,14 @@ def test_build_happy_path_with_local_container_mode(
         mock_mode.prepare.side_effect = lambda: None
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data is None
             and role == mock_role_arn
             and env == {}
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -968,13 +974,14 @@ def test_build_happy_path_with_localContainer_mode_overwritten_with_sagemaker_mo
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data is None
             and role == mock_role_arn
             and env == {}
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 
@@ -1119,13 +1126,14 @@ def test_build_happy_path_with_sagemaker_endpoint_mode_overwritten_with_local_co
         )
 
         mock_model_obj = Mock()
-        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls: (  # noqa E501
+        mock_sdk_model.side_effect = lambda image_uri, image_config, vpc_config, model_data, role, env, sagemaker_session, predictor_cls, name: (  # noqa E501
             mock_model_obj
             if image_uri == mock_image_uri
             and model_data == model_data
             and role == mock_role_arn
             and env == ENV_VARS
             and sagemaker_session == mock_session
+            and "model-name-" in name
             else None
         )
 

tests/unit/sagemaker/serve/builder/test_tei_builder.py

@@ -79,6 +79,7 @@ def test_tei_builder_sagemaker_endpoint_mode_no_s3_upload_success(
         # verify SAGEMAKER_ENDPOINT deploy
         builder = ModelBuilder(
             model=MOCK_MODEL_ID,
+            name="mock_model_name",
             schema_builder=MOCK_SCHEMA_BUILDER,
             mode=Mode.SAGEMAKER_ENDPOINT,
             model_metadata={
@@ -88,7 +89,10 @@ def test_tei_builder_sagemaker_endpoint_mode_no_s3_upload_success(
 
         builder._prepare_for_mode = MagicMock()
         builder._prepare_for_mode.return_value = (None, {})
+
         model = builder.build()
+        assert model.name == "mock_model_name"
+
         builder.serve_settings.telemetry_opt_out = True
         builder._original_deploy = MagicMock()
 

tests/unit/sagemaker/serve/builder/test_tensorflow_serving_builder.py

@@ -33,6 +33,7 @@ def setUp(self):
         self.instance.image_config = {}
         self.instance.vpc_config = {}
         self.instance.modes = {}
+        self.instance.name = "model-name-mock-uuid-hex"
 
     @patch("os.makedirs")
     @patch("os.path.exists")
@@ -71,5 +72,6 @@ def test_create_tensorflow_model(self, mock_model):
             env=self.instance.env_vars,
             sagemaker_session=self.instance.sagemaker_session,
             predictor_cls=self.instance._get_tensorflow_predictor,
+            name="model-name-mock-uuid-hex",
         )
         self.assertEqual(model, mock_model.return_value)

tests/unit/sagemaker/serve/builder/test_tgi_builder.py

@@ -61,13 +61,17 @@ def test_tgi_builder_sagemaker_endpoint_mode_no_s3_upload_success(
         # verify SAGEMAKER_ENDPOINT deploy
         builder = ModelBuilder(
             model=MOCK_MODEL_ID,
+            name="mock_model_name",
             schema_builder=MOCK_SCHEMA_BUILDER,
             mode=Mode.SAGEMAKER_ENDPOINT,
         )
 
         builder._prepare_for_mode = MagicMock()
         builder._prepare_for_mode.return_value = (None, {})
+
         model = builder.build()
+        assert model.name == "mock_model_name"
+
         builder.serve_settings.telemetry_opt_out = True
         builder._original_deploy = MagicMock()
 
@@ -187,3 +191,75 @@ def test_tgi_builder_optimized_sagemaker_endpoint_mode_no_s3_upload_success(
 
         # verify that if optimized, no s3 upload occurs
         builder._prepare_for_mode.assert_called_with()
+
+    @patch(
+        "sagemaker.serve.builder.tgi_builder._get_nb_instance",
+        return_value="ml.g5.24xlarge",
+    )
+    @patch("sagemaker.serve.builder.tgi_builder._capture_telemetry", side_effect=None)
+    @patch(
+        "sagemaker.serve.builder.model_builder.get_huggingface_model_metadata",
+        return_value={"pipeline_tag": "text-generation"},
+    )
+    @patch(
+        "sagemaker.serve.builder.tgi_builder._get_model_config_properties_from_hf",
+        return_value=({}, None),
+    )
+    @patch(
+        "sagemaker.serve.builder.tgi_builder._get_default_tgi_configurations",
+        return_value=({}, None),
+    )
+    @patch(
+        "sagemaker.serve.builder.tgi_builder._get_admissible_tensor_parallel_degrees",
+        return_value=[4, 8],
+    )
+    @patch("sagemaker.serve.builder.tgi_builder._get_admissible_dtypes", return_value=["fp16"])
+    @patch("sagemaker.serve.builder.tgi_builder.datetime")
+    @patch("sagemaker.serve.builder.tgi_builder.timedelta", return_value=1800)
+    @patch("sagemaker.serve.builder.tgi_builder._serial_benchmark")
+    @patch("sagemaker.serve.builder.tgi_builder._concurrent_benchmark")
+    def test_tgi_builder_tune_success(
+        self,
+        mock_concurrent_benchmark,
+        mock_serial_benchmark,
+        mock_timedelta,
+        mock_datetime,
+        mock_get_admissible_dtypes,
+        mock_get_admissible_tensor_parallel_degrees,
+        mock_default_tgi_configurations,
+        mock_hf_model_config,
+        mock_hf_model_md,
+        mock_get_nb_instance,
+        mock_telemetry,
+    ):
+        # WHERE
+        mock_datetime.now.side_effect = [0, 100, 200]
+        mock_serial_benchmark.side_effect = [(1000, 10000, 10), (500, 5000, 50)]
+        mock_concurrent_benchmark.side_effect = [(10, 10), (50, 5)]
+
+        builder = ModelBuilder(
+            model=MOCK_MODEL_ID,
+            schema_builder=MOCK_SCHEMA_BUILDER,
+            mode=Mode.LOCAL_CONTAINER,
+            model_path=MOCK_MODEL_PATH,
+        )
+        builder._prepare_for_mode = MagicMock()
+        builder._prepare_for_mode.side_effect = None
+
+        model = builder.build()
+
+        builder.serve_settings.telemetry_opt_out = True
+        builder.modes[str(Mode.LOCAL_CONTAINER)] = MagicMock()
+        builder.pysdk_model = MagicMock()
+
+        # WHEN
+        ret_new_model = model.tune(max_tuning_duration=1800)
+
+        # THEN
+        assert ret_new_model != model
+        assert len(mock_datetime.now.call_args_list) == 3
+        assert len(mock_serial_benchmark.call_args_list) == 2
+        assert len(mock_concurrent_benchmark.call_args_list) == 2
+        assert ret_new_model.env["NUM_SHARD"] == "8"
+        assert ret_new_model.env["DTYPE"] == "fp16"
+        assert ret_new_model.env["SHARDED"] == "true"