Move TF building to an actual build() method (#23760)

* A fun new PR where I break the entire codebase again

* A fun new PR where I break the entire codebase again

* Handle cross-attention

* Move calls to model(model.dummy_inputs) to the new build() method

* Seeing what fails with the build context thing

* make fix-copies

* Let's see what fails with new build methods

* Fix the pytorch crossload build calls

* Fix the overridden build methods in vision_text_dual_encoder

* Make sure all our build methods set self.built or call super().build(), which also sets it

* make fix-copies

* Remove finished TODO

* Tentatively remove unneeded (?) line

* Transpose b in deberta correctly and remove unused threading local

* Get rid of build_with_dummies and all it stands for

* Rollback some changes to TF-PT crossloading

* Correctly call super().build()

tests/models/bart/test_modeling_tf_bart.py

@@ -328,7 +328,7 @@ class TFBartModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, PipelineTester
             old_total_size = config.vocab_size
             new_total_size = old_total_size + new_tokens_size
             model = model_class(config=copy.deepcopy(config))  # `resize_token_embeddings` mutates `config`
-            model(model.dummy_inputs)  # builds the embeddings layer
+            model.build()
             model.resize_token_embeddings(new_total_size)
 
             # fetch the output for an input exclusively made of new members of the vocabulary

tests/models/encoder_decoder/test_modeling_tf_encoder_decoder.py

@@ -1070,9 +1070,9 @@ class TFEncoderDecoderModelSaveLoadTests(unittest.TestCase):
 
         # create two random BERT models for bert2bert & initialize weights (+cross_attention weights)
         encoder = TFBertModel(config.encoder)
-        encoder(encoder.dummy_inputs)
+        encoder.build()
         decoder = TFBertLMHeadModel(config.decoder)
-        decoder(decoder.dummy_inputs)
+        decoder.build()
 
         encoder_decoder_orig = TFEncoderDecoderModel(encoder=encoder, decoder=decoder)
 

tests/models/gpt2/test_modeling_tf_gpt2.py

@@ -463,7 +463,7 @@ class TFGPT2ModelTest(TFModelTesterMixin, TFCoreModelTesterMixin, PipelineTester
                 continue
 
             model = model_class(config)
-            model(model.dummy_inputs)
+            model.build()
 
             onnx_model_proto, _ = tf2onnx.convert.from_keras(model, opset=self.onnx_min_opset)
 

tests/models/opt/test_modeling_tf_opt.py

@@ -194,7 +194,7 @@ class TFOPTModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
             else:
                 # Here we build the word embeddings weights if not exists.
                 # And then we retry to get the attribute once built.
-                model(model.dummy_inputs)
+                model.build()
                 if hasattr(embedding_layer, "weight"):
                     return embedding_layer.weight
                 else:

tests/models/vision_encoder_decoder/test_modeling_tf_vision_encoder_decoder.py

@@ -729,9 +729,9 @@ class TFVisionEncoderDecoderModelSaveLoadTests(unittest.TestCase):
 
         # create two random ViT/GPT2 models for vit-gpt2 & initialize weights (+cross_attention weights)
         encoder = TFViTModel(config.encoder)
-        encoder(encoder.dummy_inputs)
+        encoder.build()
         decoder = TFGPT2LMHeadModel(config.decoder)
-        decoder(decoder.dummy_inputs)
+        decoder.build()
 
         encoder_decoder_orig = TFVisionEncoderDecoderModel(encoder=encoder, decoder=decoder)
 

tests/models/whisper/test_modeling_tf_whisper.py

@@ -281,7 +281,7 @@ class TFWhisperModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestC
         for model_class in self.all_model_classes:
             model = model_class(config)
 
-            model(model.dummy_inputs)
+            model.build()
 
             with tempfile.TemporaryDirectory() as tmpdirname:
                 model.save_pretrained(tmpdirname, saved_model=False)