Remove bad test skips (#41109)

* remove bad skips

* remove more

* fix inits

src/transformers/models/d_fine/modeling_d_fine.py

@@ -459,6 +459,12 @@ class DFinePreTrainedModel(PreTrainedModel):
                     nn.init.constant_(layer.layers[-1].weight, 0)
                     nn.init.constant_(layer.layers[-1].bias, 0)
 
+            if hasattr(module, "reg_scale"):
+                module.reg_scale.fill_(self.config.reg_scale)
+
+            if hasattr(module, "up"):
+                module.up.fill_(self.config.up)
+
         if isinstance(module, DFineMultiscaleDeformableAttention):
             nn.init.constant_(module.sampling_offsets.weight.data, 0.0)
             default_dtype = torch.get_default_dtype()
@@ -496,6 +502,10 @@ class DFinePreTrainedModel(PreTrainedModel):
             init.constant_(module.reg_conf.layers[-1].bias, 0)
             init.constant_(module.reg_conf.layers[-1].weight, 0)
 
+        if isinstance(module, nn.LayerNorm):
+            module.weight.data.fill_(1.0)
+            module.bias.data.zero_()
+
         if hasattr(module, "weight_embedding") and self.config.learn_initial_query:
             nn.init.xavier_uniform_(module.weight_embedding.weight)
         if hasattr(module, "denoising_class_embed") and self.config.num_denoising > 0:

src/transformers/models/d_fine/modular_d_fine.py

@@ -635,6 +635,12 @@ class DFinePreTrainedModel(RTDetrPreTrainedModel):
                     nn.init.constant_(layer.layers[-1].weight, 0)
                     nn.init.constant_(layer.layers[-1].bias, 0)
 
+            if hasattr(module, "reg_scale"):
+                module.reg_scale.fill_(self.config.reg_scale)
+
+            if hasattr(module, "up"):
+                module.up.fill_(self.config.up)
+
         if isinstance(module, DFineMultiscaleDeformableAttention):
             nn.init.constant_(module.sampling_offsets.weight.data, 0.0)
             default_dtype = torch.get_default_dtype()
@@ -672,6 +678,10 @@ class DFinePreTrainedModel(RTDetrPreTrainedModel):
             init.constant_(module.reg_conf.layers[-1].bias, 0)
             init.constant_(module.reg_conf.layers[-1].weight, 0)
 
+        if isinstance(module, nn.LayerNorm):
+            module.weight.data.fill_(1.0)
+            module.bias.data.zero_()
+
         if hasattr(module, "weight_embedding") and self.config.learn_initial_query:
             nn.init.xavier_uniform_(module.weight_embedding.weight)
         if hasattr(module, "denoising_class_embed") and self.config.num_denoising > 0:

src/transformers/models/xcodec/modeling_xcodec.py

@@ -332,7 +332,6 @@ class XcodecPreTrainedModel(PreTrainedAudioTokenizerBase):
             module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
             if module.bias is not None:
                 module.bias.data.zero_()
-
         elif isinstance(module, (nn.LayerNorm, nn.GroupNorm)):
             module.bias.data.zero_()
             module.weight.data.fill_(1.0)
@@ -341,6 +340,23 @@ class XcodecPreTrainedModel(PreTrainedAudioTokenizerBase):
             if module.bias is not None:
                 k = math.sqrt(module.groups / (module.in_channels * module.kernel_size[0]))
                 nn.init.uniform_(module.bias, a=-k, b=k)
+        elif module.__class__.__name__ == "Snake1d":
+            module.alpha.data.fill_(1.0)
+        elif isinstance(module, nn.ConvTranspose1d):
+            module.reset_parameters()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=0.02)
+        elif isinstance(module, XcodecModel):
+            # The conv1d are not handled correctly, as `self.acoustic_encoder/decoder` are initialized from a PreTrainedModel,
+            # but then only the submodules are used (which are not PreTrainedModels...) -> here we reinit them as in DacModel
+            for submodule in module.acoustic_encoder.modules():
+                if isinstance(submodule, nn.Conv1d):
+                    nn.init.trunc_normal_(submodule.weight, std=0.02)
+                    nn.init.constant_(submodule.bias, 0)
+            for submodule in module.acoustic_decoder.modules():
+                if isinstance(submodule, nn.Conv1d):
+                    nn.init.trunc_normal_(submodule.weight, std=0.02)
+                    nn.init.constant_(submodule.bias, 0)
 
     def apply_weight_norm(self):
         """Apply weight norm in the acoustic encoder and decoder because the original checkpoint has weight norm applied."""
@@ -396,6 +412,9 @@ class XcodecModel(XcodecPreTrainedModel):
         self.fc2 = nn.Linear(config.hidden_size, config.acoustic_model_config.hidden_size)
         self.quantizer = XcodecResidualVectorQuantization(config)
 
+        # Initialize weights and apply final processing
+        self.post_init()
+
     @staticmethod
     def _adjust_dac_decoder(decoder: nn.Module):
         r"""

tests/models/d_fine/test_modeling_d_fine.py

@@ -361,10 +361,6 @@ class DFineModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     def test_resize_tokens_embeddings(self):
         pass
 
-    @unittest.skip(reason="Not relevant for the model")
-    def test_can_init_all_missing_weights(self):
-        pass
-
     @unittest.skip(reason="Feed forward chunking is not implemented")
     def test_feed_forward_chunking(self):
         pass

tests/models/gemma3n/test_modeling_gemma3n.py

@@ -148,7 +148,6 @@ class Gemma3nAudioModelTest(ModelTesterMixin, unittest.TestCase):
     _is_stateful = True
     main_input_name = "audio_mel"
     test_initialization = False
-    test_can_init_all_missing_weights = False
 
     def setUp(self):
         self.model_tester = Gemma3nAudioModelTester(self)

tests/models/hgnet_v2/test_modeling_hgnet_v2.py

@@ -189,10 +189,6 @@ class HGNetV2ForImageClassificationTest(ModelTesterMixin, PipelineTesterMixin, u
     def setUp(self):
         self.model_tester = HGNetV2ModelTester(self)
 
-    @unittest.skip(reason="Does not work on the tiny model.")
-    def test_model_parallelism(self):
-        super().test_model_parallelism()
-
     @unittest.skip(reason="HGNetV2 does not output attentions")
     def test_attention_outputs(self):
         pass
@@ -209,14 +205,6 @@ class HGNetV2ForImageClassificationTest(ModelTesterMixin, PipelineTesterMixin, u
     def test_model_common_attributes(self):
         pass
 
-    @unittest.skip(reason="HGNetV2 does not have a model")
-    def test_model(self):
-        pass
-
-    @unittest.skip(reason="Not relevant for the model")
-    def test_can_init_all_missing_weights(self):
-        pass
-
     def test_backbone(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_backbone(*config_and_inputs)

tests/models/xcodec/test_modeling_xcodec.py

@@ -114,7 +114,6 @@ class XcodecModelTest(ModelTesterMixin, unittest.TestCase):
     test_headmasking = False
     test_resize_embeddings = False
     test_torchscript = False
-    test_can_init_all_missing_weights = False
 
     def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
         # model does not support returning hidden states