Don't modify tied_weight_keys in-place

src/transformers/modeling_utils.py

@@ -1318,9 +1318,11 @@ def __init__(self, config: PreTrainedConfig, *inputs, **kwargs):
         self.warnings_issued = {}
         # Overwrite the class attribute to make it an instance attribute, so models like
         # `InstructBlipForConditionalGeneration` can dynamically update it without modifying the class attribute
-        # when a different component (e.g. language_model) is used.
+        # when a different component (e.g. language_model) is used. Same for `_tied_weights_keys` which pops/adds
+        # new keys dynamically depending on config values
         self._keep_in_fp32_modules = copy.copy(self.__class__._keep_in_fp32_modules)
         self._keep_in_fp32_modules_strict = copy.copy(self.__class__._keep_in_fp32_modules_strict)
+        self._tied_weights_keys = copy.copy(self.__class__._tied_weights_keys)
         self.dtype_plan = {}
 
         if isinstance(self._keep_in_fp32_modules, list):

tests/models/deformable_detr/test_modeling_deformable_detr.py

@@ -68,7 +68,6 @@ def __init__(
         num_feature_levels=4,
         encoder_n_points=2,
         decoder_n_points=6,
-        tie_word_embeddings=False,
     ):
         self.parent = parent
         self.batch_size = batch_size
@@ -89,7 +88,6 @@ def __init__(
         self.num_feature_levels = num_feature_levels
         self.encoder_n_points = encoder_n_points
         self.decoder_n_points = decoder_n_points
-        self.tie_word_embeddings = tie_word_embeddings
 
         # we also set the expected seq length for both encoder and decoder
         self.encoder_seq_length = (
@@ -151,9 +149,6 @@ def get_config(self):
             backbone=None,
             backbone_config=resnet_config,
             use_pretrained_backbone=False,
-            # FIXME; cls attr `toed_weihgt_keys` must not be modified in __init__
-            # Several models affected so for now just let it be and fix in separate PR
-            tie_word_embeddings=self.tie_word_embeddings,
         )
 
     def prepare_config_and_inputs_for_common(self):
@@ -248,6 +243,25 @@ def test_deformable_detr_object_detection_head_model(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_deformable_detr_object_detection_head_model(*config_and_inputs)
 
+    def test_tie_weights_is_not_modified(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.tie_word_embeddings = True
+
+        config.with_box_refine = True
+        config.two_stage = True
+
+        model = DeformableDetrForObjectDetection(config)
+        self.assertTrue("model.decoder.bbox_embed" in model._tied_weights_keys)
+        self.assertTrue("model.decoder.class_embed" in model._tied_weights_keys)
+
+        # if we update config attr, model's tied weights keys also change
+        config.with_box_refine = False
+        config.two_stage = False
+
+        model = DeformableDetrForObjectDetection(config)
+        self.assertFalse("model.decoder.bbox_embed" in model._tied_weights_keys)
+        self.assertFalse("model.decoder.class_embed" in model._tied_weights_keys)
+
     @unittest.skip(reason="Deformable DETR does not use inputs_embeds")
     def test_inputs_embeds(self):
         pass

tests/models/grounding_dino/test_modeling_grounding_dino.py

@@ -322,6 +322,19 @@ def test_feed_forward_chunking(self):
     def test_load_save_without_tied_weights(self):
         pass
 
+    def test_tie_weights_is_not_modified(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.tie_word_embeddings = True
+
+        config.decoder_bbox_embed_share = False
+        model = GroundingDinoForObjectDetection(config)
+        self.assertFalse(r"bbox_embed.(?![0])\d+" in model._tied_weights_keys)
+
+        # if we update config attr, model's tied weights keys also change
+        config.decoder_bbox_embed_share = True
+        model = GroundingDinoForObjectDetection(config)
+        self.assertTrue(r"bbox_embed.(?![0])\d+" in model._tied_weights_keys)
+
     def test_attention_outputs(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         config.return_dict = True

tests/models/mm_grounding_dino/test_modeling_mm_grounding_dino.py

@@ -327,6 +327,11 @@ def test_feed_forward_chunking(self):
     def test_load_save_without_tied_weights(self):
         pass
 
+    # Ignore copy
+    def test_tie_weights_is_not_modified(self):
+        # this model doesn't need a test
+        pass
+
     def test_attention_outputs(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
         config.return_dict = True