Deprecate dtype per sub config

src/transformers/modeling_utils.py

@@ -759,7 +759,6 @@ def _get_dtype(
     2. Else, use the dtype provided as a dict or str
     """
     is_sharded = sharded_metadata is not None
-    asked_dtype = dtype
 
     if dtype is not None:
         if isinstance(dtype, str):
@@ -806,24 +805,21 @@ def _get_dtype(
     if isinstance(dtype, dict):
         main_dtype = dtype.get("", torch.get_default_dtype())
         main_dtype = getattr(torch, main_dtype) if isinstance(main_dtype, str) else main_dtype
+
+        logger.warning_once(
+            "Using different dtypes per module is deprecated and will be removed in future versions "
+            "Setting different dtypes per backbone model might cause device errors downstream, therefore "
+            f"setting the dtype={main_dtype} for all modules."
+        )
+
     else:
         main_dtype = dtype
 
     # Set it on the config and subconfigs
     config.dtype = main_dtype
     for sub_config_key in config.sub_configs:
         if (sub_config := getattr(config, sub_config_key)) is not None:
-            # The dtype was "auto" -> try to read the subconfig dtype value if any
-            if asked_dtype == "auto":
-                sub_dtype = getattr(sub_config, "dtype", main_dtype)
-                sub_dtype = getattr(torch, sub_dtype) if isinstance(sub_dtype, str) else sub_dtype
-            # The dtype was provided as a dict, try to see if we match the subconfig name
-            elif isinstance(dtype, dict):
-                sub_dtype = dtype.get(sub_config_key, main_dtype)
-                sub_dtype = getattr(torch, sub_dtype) if isinstance(sub_dtype, str) else sub_dtype
-            else:
-                sub_dtype = main_dtype
-            sub_config.dtype = sub_dtype
+            sub_config.dtype = main_dtype
 
     return config, main_dtype
 

tests/utils/test_modeling_utils.py

@@ -569,6 +569,8 @@ def test_model_from_config_dtype_composite(self):
         """
         Test that from_pretrained works with dtype being as a dict per each sub-config in composite config
         Tiny-Llava has saved auto dtype as `torch.float32` for all modules.
+        Note, this is a deprecated feature and we fallback to main dtype in all cases below. This test checks
+        if the dtype fallback works correctly.
         """
         # Load without dtype specified
         model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA)
@@ -587,42 +589,42 @@ def test_model_from_config_dtype_composite(self):
         self.assertEqual(model.model.vision_tower.dtype, torch.float16)
         self.assertIsInstance(model.config.dtype, torch.dtype)
 
-        # should be able to set dtype as a dict for each sub-config
+        # should be able to accept dtype as a dict for each sub-config
         model = LlavaForConditionalGeneration.from_pretrained(
             TINY_LLAVA, dtype={"text_config": "float32", "vision_config": "float16", "": "bfloat16"}
         )
-        self.assertEqual(model.model.language_model.dtype, torch.float32)
-        self.assertEqual(model.model.vision_tower.dtype, torch.float16)
+        self.assertEqual(model.model.language_model.dtype, torch.bfloat16)
+        self.assertEqual(model.model.vision_tower.dtype, torch.bfloat16)
         self.assertEqual(model.model.multi_modal_projector.linear_1.weight.dtype, torch.bfloat16)
         self.assertIsInstance(model.config.dtype, torch.dtype)
 
-        # should be able to set the values as torch.dtype (not str)
+        # should be able to accept the values as torch.dtype (not str)
         model = LlavaForConditionalGeneration.from_pretrained(
             TINY_LLAVA, dtype={"text_config": torch.float32, "vision_config": torch.float16, "": torch.bfloat16}
         )
-        self.assertEqual(model.model.language_model.dtype, torch.float32)
-        self.assertEqual(model.model.vision_tower.dtype, torch.float16)
+        self.assertEqual(model.model.language_model.dtype, torch.bfloat16)
+        self.assertEqual(model.model.vision_tower.dtype, torch.bfloat16)
         self.assertEqual(model.model.multi_modal_projector.linear_1.weight.dtype, torch.bfloat16)
         self.assertIsInstance(model.config.dtype, torch.dtype)
 
-        # should be able to set the values in configs directly and pass it to `from_pretrained`
+        # should be able to accept the values in configs directly and pass it to `from_pretrained`
         config = copy.deepcopy(model.config)
         config.text_config.dtype = torch.float32
         config.vision_config.dtype = torch.bfloat16
         config.dtype = torch.float16
         model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA, config=config, dtype="auto")
-        self.assertEqual(model.model.language_model.dtype, torch.float32)
-        self.assertEqual(model.model.vision_tower.dtype, torch.bfloat16)
+        self.assertEqual(model.model.language_model.dtype, torch.float16)
+        self.assertEqual(model.model.vision_tower.dtype, torch.float16)
         self.assertEqual(model.model.multi_modal_projector.linear_1.weight.dtype, torch.float16)
         self.assertIsInstance(model.config.dtype, torch.dtype)
 
         # but if the model has `_keep_in_fp32_modules` then those modules should be in fp32 no matter what
         LlavaForConditionalGeneration._keep_in_fp32_modules = ["multi_modal_projector"]
         model = LlavaForConditionalGeneration.from_pretrained(TINY_LLAVA, config=config, dtype="auto")
         self.assertEqual(
-            model.model.language_model.dtype, torch.float32
+            model.model.language_model.dtype, torch.float16
         )  # remember config says float32 for text_config
-        self.assertEqual(model.model.vision_tower.dtype, torch.bfloat16)
+        self.assertEqual(model.model.vision_tower.dtype, torch.float16)
         self.assertEqual(model.model.multi_modal_projector.linear_1.weight.dtype, torch.float32)
         self.assertIsInstance(model.config.dtype, torch.dtype)