Deal with nested configs better in base class (#25237)

* Deal better with nested configs

* Fixes

* More fixes

* Fix last test

* Clean up existing configs

* Remove hack in MPT Config

* Update src/transformers/configuration_utils.py

Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

* Fix setting a nested config via dict in the kwargs

* Adapt common test

* Add test for nested config load with dict

---------

Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

tests/test_configuration_common.py

@@ -118,9 +118,11 @@ class ConfigTester(object):
 
     def check_config_can_be_init_without_params(self):
         if self.config_class.is_composition:
-            return
-        config = self.config_class()
-        self.parent.assertIsNotNone(config)
+            with self.parent.assertRaises(ValueError):
+                config = self.config_class()
+        else:
+            config = self.config_class()
+            self.parent.assertIsNotNone(config)
 
     def check_config_arguments_init(self):
         kwargs = copy.deepcopy(config_common_kwargs)

tests/test_configuration_utils.py

@@ -210,6 +210,13 @@ class ConfigTestUtils(unittest.TestCase):
                 f" {', '.join(keys_with_defaults)}."
             )
 
+    def test_nested_config_load_from_dict(self):
+        config = AutoConfig.from_pretrained(
+            "hf-internal-testing/tiny-random-CLIPModel", text_config={"num_hidden_layers": 2}
+        )
+        self.assertNotIsInstance(config.text_config, dict)
+        self.assertEqual(config.text_config.__class__.__name__, "CLIPTextConfig")
+
     def test_from_pretrained_subfolder(self):
         with self.assertRaises(OSError):
             # config is in subfolder, the following should not work without specifying the subfolder