🚨🚨🚨 Enforce single model initialization (#21431)

* Enforce single model initialization

* Add OneFormer example for problem 3

* Do it the Stas way

* Actually rename the uses...

* Rewrite test

* Try to change the test this way

* Fix all init slow/fast tests

* Break connection

* Fix more tests

* Fix test for initialization

* Remove custom test

* Quality

* Fix last failing tests

* The end?

docs/source/en/add_new_model.mdx

@@ -492,6 +492,48 @@ model = BrandNewBertModel(BrandNewBertConfig())
 The above command will create a model according to the default parameters as defined in `BrandNewBertConfig()` with
 random weights, thus making sure that the `init()` methods of all components works.
 
+Note that all random initialization should happen in the `_init_weights` method of your `BrandnewBertPreTrainedModel`
+class. It should initialize all leaf modules depending on the variables of the config. Here is an example with the
+BERT `_init_weights` method:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstance(module, nn.Linear):
+        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.Embedding):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.padding_idx is not None:
+            module.weight.data[module.padding_idx].zero_()
+    elif isinstance(module, nn.LayerNorm):
+        module.bias.data.zero_()
+        module.weight.data.fill_(1.0)
+```
+
+You can have some more custom schemes if you need a special initialization for some modules. For instance, in
+`Wav2Vec2ForPreTraining`, the last two linear layers need to have the initialization of the regular PyTorch `nn.Linear`
+but all the other ones should use an initialization as above. This is coded like this:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstnace(module, Wav2Vec2ForPreTraining):
+        module.project_hid.reset_parameters()
+        module.project_q.reset_parameters()
+        module.project_hid._is_hf_initialized = True
+        module.project_q._is_hf_initialized = True
+    elif isinstance(module, nn.Linear):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.bias is not None:
+            module.bias.data.zero_()
+```
+
+The `_is_hf_initialized` flag is internally used to make sure we only initialize a submodule once. By setting it to
+`True` for `module.project_q` and `module.project_hid`, we make sure the custom initialization we did is not overridden later on,
+the `_init_weights` function won't be applied to them.
+
 **6. Write a conversion script**
 
 Next, you should write a conversion script that lets you convert the checkpoint you used to debug *brand_new_bert* in

src/transformers/modeling_utils.py

@@ -436,6 +436,17 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
             )
 
 
+def set_initialized_submodules(model, state_dict_keys):
+    """
+    Sets the `_is_hf_initialized` flag in all submodules of a given model when all its weights are in the loaded state
+    dict.
+    """
+    for module_name, module in model.named_modules():
+        loaded_keys = [k.replace(f"{module_name}.", "") for k in state_dict_keys if k.startswith(f"{module_name}.")]
+        if len(set(module.state_dict().keys()) - set(loaded_keys)) == 0:
+            module._is_hf_initialized = True
+
+
 def _load_state_dict_into_model(model_to_load, state_dict, start_prefix):
     # Convert old format to new format if needed from a PyTorch state_dict
     old_keys = []
@@ -1176,7 +1187,16 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         """
         Initialize the weights. This method should be overridden by derived class.
         """
-        raise NotImplementedError(f"Make sure `_init_weights` is implemented for {self.__class__}")
+        pass
+
+    def _initialize_weights(self, module):
+        """
+        Initialize the weights if they are not already initialized.
+        """
+        if getattr(module, "_is_hf_initialized", False):
+            return
+        self._init_weights(module)
+        module._is_hf_initialized = True
 
     def tie_weights(self):
         """
@@ -1505,7 +1525,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
     def init_weights(self):
         """
-        If needed prunes and maybe initializes weights.
+        If needed prunes and maybe initializes weights. If using a custom `PreTrainedModel`, you need to implement any
+        initialization logic in `_init_weights`.
         """
         # Prune heads if needed
         if self.config.pruned_heads:
@@ -1513,7 +1534,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
         if _init_weights:
             # Initialize weights
-            self.apply(self._init_weights)
+            self.apply(self._initialize_weights)
 
             # Tie weights should be skipped when not initializing all weights
             # since from_pretrained(...) calls tie weights anyways
@@ -2713,11 +2734,15 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
         # retrieve unintialized modules and initialize before maybe overriding that with the pretrained weights.
         if _fast_init:
-            uninitialized_modules = model.retrieve_modules_from_names(
-                missing_keys, add_prefix=add_prefix_to_model, remove_prefix=remove_prefix_from_model
-            )
-            for module in uninitialized_modules:
-                model._init_weights(module)
+            if remove_prefix_from_model:
+                _loaded_keys = [f"{prefix}.{k}" for k in loaded_keys]
+            elif add_prefix_to_model:
+                _loaded_keys = [k[len(prefix) + 1 :] for k in loaded_keys]
+            else:
+                _loaded_keys = loaded_keys
+            set_initialized_submodules(model, _loaded_keys)
+            # This will only initialize submodules that are not marked as initialized by the line above.
+            model.apply(model._initialize_weights)
 
         # Set some modules to fp32 if any
         if keep_in_fp32_modules is not None:

src/transformers/models/altclip/modeling_altclip.py

@@ -1067,10 +1067,12 @@ class AltCLIPPreTrainedModel(PreTrainedModel):
                 module.text_projection.weight,
                 std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
             )
+            module.text_projection._is_hf_initialized = True
             nn.init.normal_(
                 module.visual_projection.weight,
                 std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
             )
+            module.visual_projection._is_hf_initialized = True
         elif isinstance(module, nn.LayerNorm):
             module.bias.data.zero_()
             module.weight.data.fill_(1.0)

src/transformers/models/bart/modeling_bart.py

@@ -1473,8 +1473,9 @@ class BartForSequenceClassification(BartPretrainedModel):
             config.num_labels,
             config.classifier_dropout,
         )
-        self.model._init_weights(self.classification_head.dense)
-        self.model._init_weights(self.classification_head.out_proj)
+
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -1601,7 +1602,8 @@ class BartForQuestionAnswering(BartPretrainedModel):
         self.model = BartModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.model._init_weights(self.qa_outputs)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(BART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -2658,8 +2658,9 @@ class BigBirdPegasusForSequenceClassification(BigBirdPegasusPreTrainedModel):
             config.num_labels,
             config.classifier_dropout,
         )
-        self.model._init_weights(self.classification_head.dense)
-        self.model._init_weights(self.classification_head.out_proj)
+
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -2785,7 +2786,8 @@ class BigBirdPegasusForQuestionAnswering(BigBirdPegasusPreTrainedModel):
         self.model = BigBirdPegasusModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.model._init_weights(self.qa_outputs)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(BIGBIRD_PEGASUS_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/models/fsmt/modeling_fsmt.py

@@ -1186,6 +1186,9 @@ class FSMTForConditionalGeneration(PretrainedFSMTModel):
         base_model = FSMTModel(config)
         self.model = base_model
 
+        # Initialize weights and apply final processing
+        self.post_init()
+
     @add_start_docstrings_to_model_forward(FSMT_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
     @add_end_docstrings(FSMT_GENERATION_EXAMPLE)

src/transformers/models/led/modeling_led.py

@@ -2543,8 +2543,9 @@ class LEDForSequenceClassification(LEDPreTrainedModel):
             config.num_labels,
             config.classifier_dropout,
         )
-        self.led._init_weights(self.classification_head.dense)
-        self.led._init_weights(self.classification_head.out_proj)
+
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(LED_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -2672,7 +2673,8 @@ class LEDForQuestionAnswering(LEDPreTrainedModel):
         self.led = LEDModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.led._init_weights(self.qa_outputs)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(LED_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/models/maskformer/modeling_maskformer_swin.py

@@ -866,6 +866,9 @@ class MaskFormerSwinBackbone(MaskFormerSwinPreTrainedModel, BackboneMixin):
 
         self.hidden_states_norms = nn.ModuleList([nn.LayerNorm(num_channels) for num_channels in self.channels])
 
+        # Initialize weights and apply final processing
+        self.post_init()
+
     @property
     def channels(self):
         return [self.out_feature_channels[name] for name in self.out_features]

src/transformers/models/mbart/modeling_mbart.py

@@ -1447,8 +1447,9 @@ class MBartForSequenceClassification(MBartPreTrainedModel):
             config.num_labels,
             config.classifier_dropout,
         )
-        self.model._init_weights(self.classification_head.dense)
-        self.model._init_weights(self.classification_head.out_proj)
+
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
@@ -1574,7 +1575,8 @@ class MBartForQuestionAnswering(MBartPreTrainedModel):
         self.model = MBartModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.model._init_weights(self.qa_outputs)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(MBART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/models/mvp/modeling_mvp.py

@@ -1610,8 +1610,8 @@ class MvpForSequenceClassification(MvpPreTrainedModel):
             config.classifier_dropout,
         )
 
-        self.model._init_weights(self.classification_head.dense)
-        self.model._init_weights(self.classification_head.out_proj)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     def set_lightweight_tuning(self):
         self.model.set_lightweight_tuning()
@@ -1737,7 +1737,8 @@ class MvpForQuestionAnswering(MvpPreTrainedModel):
         self.model = MvpModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.model._init_weights(self.qa_outputs)
+        # Initialize weights and apply final processing
+        self.post_init()
 
     def set_lightweight_tuning(self):
         self.model.set_lightweight_tuning()

src/transformers/models/oneformer/modeling_oneformer.py

@@ -2801,6 +2801,7 @@ class OneFormerPreTrainedModel(PreTrainedModel):
         elif isinstance(module, OneFormerTransformerDecoder):
             nn.init.xavier_uniform_(module.query_input_projection.weight, gain=xavier_std)
             nn.init.constant_(module.query_input_projection.bias, 0)
+            module.query_input_projection._is_hf_initialized = True
         elif isinstance(module, OneFormerPixelDecoderEncoderMultiscaleDeformableAttention):
             nn.init.constant_(module.sampling_offsets.weight.data, 0.0)
             thetas = torch.arange(module.n_heads, dtype=torch.float32) * (2.0 * math.pi / module.n_heads)

src/transformers/models/plbart/modeling_plbart.py

@@ -1420,8 +1420,9 @@ class PLBartForSequenceClassification(PLBartPreTrainedModel):
             config.num_labels,
             config.classifier_dropout,
         )
-        self.model._init_weights(self.classification_head.dense)
-        self.model._init_weights(self.classification_head.out_proj)
+
+        # Initialize weights and apply final processing
+        self.post_init()
 
     @add_start_docstrings_to_model_forward(PLBART_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(

src/transformers/models/upernet/modeling_upernet.py

@@ -301,6 +301,12 @@ class UperNetPreTrainedModel(PreTrainedModel):
     main_input_name = "pixel_values"
     supports_gradient_checkpointing = True
 
+    def _init_weights(self, module):
+        if isinstance(module, UperNetPreTrainedModel):
+            module.backbone.init_weights()
+            module.decode_head.init_weights()
+            module.auxiliary_head.init_weights()
+
     def init_weights(self):
         """Initialize the weights"""
         self.backbone.init_weights()

src/transformers/models/wav2vec2/modeling_wav2vec2.py

@@ -1049,8 +1049,14 @@ class Wav2Vec2PreTrainedModel(PreTrainedModel):
 
     def _init_weights(self, module):
         """Initialize the weights"""
+        # Wav2Vec2ForPreTraining last 2 linear layers need standard Linear init.
+        if isinstance(module, Wav2Vec2ForPreTraining):
+            module.project_hid.reset_parameters()
+            module.project_q.reset_parameters()
+            module.project_hid._is_hf_initialized = True
+            module.project_q._is_hf_initialized = True
         # gumbel softmax requires special init
-        if isinstance(module, Wav2Vec2GumbelVectorQuantizer):
+        elif isinstance(module, Wav2Vec2GumbelVectorQuantizer):
             module.weight_proj.weight.data.normal_(mean=0.0, std=1)
             module.weight_proj.bias.data.zero_()
             nn.init.uniform_(module.codevectors)
@@ -1345,13 +1351,12 @@ class Wav2Vec2ForPreTraining(Wav2Vec2PreTrainedModel):
 
         self.quantizer = Wav2Vec2GumbelVectorQuantizer(config)
 
-        # Initialize weights and apply final processing
-        self.post_init()
-
-        # make sure that project_hid & project_q are initialized like normal linear layers
         self.project_hid = nn.Linear(config.hidden_size, config.proj_codevector_dim)
         self.project_q = nn.Linear(config.codevector_dim, config.proj_codevector_dim)
 
+        # Initialize weights and apply final processing
+        self.post_init()
+
     def set_gumbel_temperature(self, temperature: int):
         """
         Set the Gumbel softmax temperature to a given value. Only necessary for training

src/transformers/models/wav2vec2_conformer/modeling_wav2vec2_conformer.py

@@ -1089,8 +1089,14 @@ class Wav2Vec2ConformerPreTrainedModel(PreTrainedModel):
 
     def _init_weights(self, module):
         """Initialize the weights"""
+        # Wav2Vec2ForPreTraining last 2 linear layers need standard Linear init.
+        if isinstance(module, Wav2Vec2ConformerForPreTraining):
+            module.project_hid.reset_parameters()
+            module.project_q.reset_parameters()
+            module.project_hid._is_hf_initialized = True
+            module.project_q._is_hf_initialized = True
         # gumbel softmax requires special init
-        if isinstance(module, Wav2Vec2ConformerGumbelVectorQuantizer):
+        elif isinstance(module, Wav2Vec2ConformerGumbelVectorQuantizer):
             module.weight_proj.weight.data.normal_(mean=0.0, std=1)
             module.weight_proj.bias.data.zero_()
             nn.init.uniform_(module.codevectors)
@@ -1381,13 +1387,12 @@ class Wav2Vec2ConformerForPreTraining(Wav2Vec2ConformerPreTrainedModel):
 
         self.quantizer = Wav2Vec2ConformerGumbelVectorQuantizer(config)
 
-        # Initialize weights and apply final processing
-        self.post_init()
-
-        # make sure that project_hid & project_q are initialized like normal linear layers
         self.project_hid = nn.Linear(config.hidden_size, config.proj_codevector_dim)
         self.project_q = nn.Linear(config.codevector_dim, config.proj_codevector_dim)
 
+        # Initialize weights and apply final processing
+        self.post_init()
+
     # Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2ForPreTraining.set_gumbel_temperature
     def set_gumbel_temperature(self, temperature: int):
         """

src/transformers/models/wavlm/modeling_wavlm.py

@@ -962,7 +962,6 @@ class WavLMAdapterLayer(nn.Module):
         return hidden_states
 
 
-# Copied from transformers.models.wav2vec2.modeling_wav2vec2.Wav2Vec2PreTrainedModel with Wav2Vec2->WavLM, wav2vec2->wavlm
 class WavLMPreTrainedModel(PreTrainedModel):
     """
     An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained

tests/models/bart/test_modeling_bart.py

@@ -1496,3 +1496,6 @@ class BartStandaloneDecoderModelTest(ModelTesterMixin, GenerationTesterMixin, un
     def test_retain_grad_hidden_states_attentions(self):
         # decoder cannot keep gradients
         return
+
+    def test_save_load_fast_init_from_base(self):
+        pass

tests/models/deta/test_modeling_deta.py

@@ -410,8 +410,13 @@ class DetaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
         configs_no_init = _config_zero_init(config)
         for model_class in self.all_model_classes:
             model = model_class(config=configs_no_init)
+            # Skip the check for the backbone
+            for name, module in model.named_modules():
+                if module.__class__.__name__ == "DetaBackboneWithPositionalEncodings":
+                    backbone_params = [f"{name}.{key}" for key in module.state_dict().keys()]
+                    break
+
             for name, param in model.named_parameters():
-                if param.requires_grad:
                 if param.requires_grad:
                     if (
                         "level_embed" in name
@@ -419,6 +424,7 @@ class DetaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
                         or "value_proj" in name
                         or "output_proj" in name
                         or "reference_points" in name
+                        or name in backbone_params
                     ):
                         continue
                     self.assertIn(

tests/models/dpt/test_modeling_dpt.py

@@ -24,7 +24,7 @@ from transformers.models.auto import get_values
 from transformers.testing_utils import require_torch, require_vision, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
 
 
 if is_torch_available():
@@ -242,6 +242,29 @@ class DPTModelTest(ModelTesterMixin, unittest.TestCase):
             loss = model(**inputs).loss
             loss.backward()
 
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            # Skip the check for the backbone
+            backbone_params = []
+            for name, module in model.named_modules():
+                if module.__class__.__name__ == "DPTViTHybridEmbeddings":
+                    backbone_params = [f"{name}.{key}" for key in module.state_dict().keys()]
+                    break
+
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if name in backbone_params:
+                        continue
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in DPT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:

tests/models/dpt/test_modeling_dpt_hybrid.py

@@ -24,7 +24,7 @@ from transformers.models.auto import get_values
 from transformers.testing_utils import require_torch, require_vision, slow, torch_device
 
 from ...test_configuration_common import ConfigTester
-from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor, ids_tensor
 
 
 if is_torch_available():
@@ -256,6 +256,29 @@ class DPTModelTest(ModelTesterMixin, unittest.TestCase):
             loss = model(**inputs).loss
             loss.backward()
 
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            # Skip the check for the backbone
+            backbone_params = []
+            for name, module in model.named_modules():
+                if module.__class__.__name__ == "DPTViTHybridEmbeddings":
+                    backbone_params = [f"{name}.{key}" for key in module.state_dict().keys()]
+                    break
+
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    if name in backbone_params:
+                        continue
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
     @slow
     def test_model_from_pretrained(self):
         for model_name in DPT_PRETRAINED_MODEL_ARCHIVE_LIST[1:]: