remove UTs for now

Author: Ashish Gupta

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

@@ -2667,40 +2667,6 @@ def test_optimize_exclusive_args(self, mock_get_serve_setting):
             ),
         )
 
-    @patch.object(ModelBuilder, "_get_serve_setting", autospec=True)
-    def test_optimize_exclusive_sharding(self, mock_get_serve_setting):
-        mock_sagemaker_session = Mock()
-        model_builder = ModelBuilder(
-            model="meta-textgeneration-llama-3-70b",
-            sagemaker_session=mock_sagemaker_session,
-        )
-
-        self.assertRaisesRegex(
-            ValueError,
-            "Sharding config is mutually exclusive and cannot be combined with any other optimization.",
-            lambda: model_builder.optimize(
-                quantization_config={"OverrideEnvironment": {"OPTION_QUANTIZE": "awq"}},
-                compilation_config={"OverrideEnvironment": {"OPTION_QUANTIZE": "awq"}},
-                sharding_config={"OverrideEnvironment": {"OPTION_QUANTIZE": "awq"}},
-            ),
-        )
-
-    @patch.object(ModelBuilder, "_get_serve_setting", autospec=True)
-    def test_optimize_exclusive_sharding_args(self, mock_get_serve_setting):
-        mock_sagemaker_session = Mock()
-        model_builder = ModelBuilder(
-            model="meta-textgeneration-llama-3-70b",
-            sagemaker_session=mock_sagemaker_session,
-        )
-
-        self.assertRaisesRegex(
-            ValueError,
-            "OPTION_TENSOR_PARALLEL_DEGREE is required environment variable with Sharding config.",
-            lambda: model_builder.optimize(
-                sharding_config={"OverrideEnvironment": {"OPTION_QUANTIZE": "awq"}},
-            ),
-        )
-
     @patch.object(ModelBuilder, "_prepare_for_mode")
     @patch.object(ModelBuilder, "_get_serve_setting", autospec=True)
     def test_optimize_for_hf_with_custom_s3_path(

tests/unit/sagemaker/serve/utils/test_optimize_utils.py

@@ -261,7 +261,7 @@ def test_is_s3_uri(s3_uri, expected):
 
 
 @pytest.mark.parametrize(
-    "quantization_config, compilation_config, sharding_config, expected_config, expected_env",
+    "quantization_config, compilation_config, expected_config, expected_env",
     [
         (
             None,
@@ -270,7 +270,6 @@ def test_is_s3_uri(s3_uri, expected):
                     "OPTION_TENSOR_PARALLEL_DEGREE": "2",
                 }
             },
-            None,
             {
                 "ModelCompilationConfig": {
                     "OverrideEnvironment": {
@@ -289,7 +288,6 @@ def test_is_s3_uri(s3_uri, expected):
                 }
             },
             None,
-            None,
             {
                 "ModelQuantizationConfig": {
                     "OverrideEnvironment": {
@@ -301,26 +299,7 @@ def test_is_s3_uri(s3_uri, expected):
                 "OPTION_TENSOR_PARALLEL_DEGREE": "2",
             },
         ),
-        (
-            None,
-            None,
-            {
-                "OverrideEnvironment": {
-                    "OPTION_TENSOR_PARALLEL_DEGREE": "2",
-                }
-            },
-            {
-                "ModelShardingConfig": {
-                    "OverrideEnvironment": {
-                        "OPTION_TENSOR_PARALLEL_DEGREE": "2",
-                    }
-                },
-            },
-            {
-                "OPTION_TENSOR_PARALLEL_DEGREE": "2",
-            },
-        ),
-        (None, None, None, None, None),
+        (None, None, None, None),
     ],
 )
 def test_extract_optimization_config_and_env(