Use deep copies instead of shallow copies for bbox_embed in GroundingDINO decoder (#37333).

src/transformers/models/grounding_dino/modeling_grounding_dino.py

@@ -14,6 +14,8 @@
 # limitations under the License.
 """PyTorch Grounding DINO model."""
 
+import copy
+
 import math
 import warnings
 from dataclasses import dataclass
@@ -2458,7 +2460,9 @@ def __init__(self, config: GroundingDinoConfig):
             shared_head = GroundingDinoMLPPredictionHead(
                 input_dim=config.d_model, hidden_dim=config.d_model, output_dim=4, num_layers=3
             )
-            self.bbox_embed = nn.ModuleList([shared_head] * config.decoder_layers)
+            #self.bbox_embed = nn.ModuleList([shared_head] * config.decoder_layers)
+
+            self.bbox_embed = nn.ModuleList( [copy.deepcopy(shared_head) for _ in range(config.decoder_layers)])
         else:
             # each layer has its own head (implicit deep copy through a new instance)
             self.bbox_embed = nn.ModuleList(

tests/models/Rabi_Testcases/GD_Layer_Independency_Testcase.py

@@ -0,0 +1,21 @@
+import pytest
+import torch
+from transformers.models.grounding_dino.modeling_grounding_dino import GroundingDinoModel
+from transformers.models.grounding_dino.configuration_grounding_dino import GroundingDinoConfig
+
+def test_bbox_embed_heads_are_independent_with_custom_config():
+    config = GroundingDinoConfig(
+        decoder_layers=2,
+        decoder_bbox_embed_share=True,
+        d_model=256,
+        num_queries=1,
+                             )
+    model = GroundingDinoModel(config)
+    assert model.bbox_embed[0] is not model.bbox_embed[1]
+    original_weight = model.bbox_embed[1].layers[0].weight.clone()
+
+    with torch.no_grad():
+        model.bbox_embed[0].layers[0].weight.add_(10.0)
+
+    assert not torch.equal(model.bbox_embed[0].layers[0].weight, original_weight)
+    assert torch.equal(model.bbox_embed[1].layers[0].weight, original_weight)

tests/models/Rabi_Testcases/GD_Unique_Instances_Testcase.py

@@ -0,0 +1,13 @@
+import pytest
+from transformers import AutoConfig
+from transformers.models.grounding_dino.modeling_grounding_dino import GroundingDinoModel
+
+def test_bbox_embed_instances_are_unique():
+    config = AutoConfig.for_model("grounding_dino")
+    config.decoder_layers = 2
+    config.decoder_bbox_embed_share = True
+
+    model = GroundingDinoModel(config)
+
+    # Ensure each layer has a unique bbox_embed head
+    assert model.bbox_embed[0] is not model.bbox_embed[1]