Fix: loading DBRX back from saved path

src/transformers/modeling_utils.py

@@ -4029,7 +4029,9 @@ def from_pretrained(
                             sub_config = getattr(config, sub_config_key)
                             sub_config.torch_dtype = torch_dtype
                 elif isinstance(torch_dtype, torch.dtype):
-                    pass
+                    for sub_config_key in config.sub_configs.keys():
+                        sub_config = getattr(config, sub_config_key)
+                        sub_config.torch_dtype = torch_dtype
                 elif isinstance(torch_dtype, dict):
                     for key, curr_dtype in torch_dtype.items():
                         if hasattr(config, key):

src/transformers/models/dbrx/configuration_dbrx.py

@@ -57,7 +57,7 @@ def __init__(
         self.kv_n_heads = kv_n_heads
         self.rope_theta = rope_theta
 
-        for k in ["model_type", "attn_implementation", "transformers_version", "_commit_hash"]:
+        for k in ["model_type", "attn_implementation", "transformers_version", "_commit_hash", "torch_dtype"]:
             if k in kwargs:
                 kwargs.pop(k)
         if len(kwargs) != 0:
@@ -109,7 +109,7 @@ def __init__(
         self.moe_loss_weight = moe_loss_weight
         self.moe_normalize_expert_weights = moe_normalize_expert_weights
 
-        for k in ["model_type", "attn_implementation", "transformers_version", "_commit_hash"]:
+        for k in ["model_type", "attn_implementation", "transformers_version", "_commit_hash", "torch_dtype"]:
             if k in kwargs:
                 kwargs.pop(k)
         if len(kwargs) != 0:

tests/test_modeling_common.py

@@ -331,6 +331,12 @@ def check_save_load(out1, out2):
                 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 @@ def test_model_from_config_torch_dtype_str(self):
     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)