Fix: loading DBRX back from saved path (#35728)

* fix dtype as dict for some models + add test

* add comment in tests

tests/test_modeling_common.py

@@ -331,6 +331,12 @@ class ModelTesterMixin:
                 with torch.no_grad():
                     second = model(**self._prepare_for_class(inputs_dict, model_class))[0]
 
+                # Save and load second time because `from_pretrained` adds a bunch of new config fields
+                # so we need to make sure those fields can be loaded back after saving
+                # Simply init as `model(config)` doesn't add those fields
+                model.save_pretrained(tmpdirname)
+                model = model_class.from_pretrained(tmpdirname)
+
             if isinstance(first, tuple) and isinstance(second, tuple):
                 for tensor1, tensor2 in zip(first, second):
                     check_save_load(tensor1, tensor2)

tests/utils/test_modeling_utils.py

@@ -466,13 +466,14 @@ class ModelUtilsTest(TestCasePlus):
     def test_model_from_config_torch_dtype_composite(self):
         """
         Test that from_pretrained works with torch_dtype being as a dict per each sub-config in composite config
+        Tiny-Llava has saved auto dtype as `torch.float32` for all modules.
         """
         # should be able to set torch_dtype as a simple string and the model loads it correctly
         model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA, torch_dtype="float32")
         self.assertEqual(model.language_model.dtype, torch.float32)
         self.assertEqual(model.vision_tower.dtype, torch.float32)
 
-        model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA, torch_dtype="float16")
+        model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA, torch_dtype=torch.float16)
         self.assertEqual(model.language_model.dtype, torch.float16)
         self.assertEqual(model.vision_tower.dtype, torch.float16)