Add TFEncoderDecoderModel + Add cross-attention to some TF models (#13222)

* Add cross attentions to TFGPT2Model

* Add TFEncoderDecoderModel

* Add TFBaseModelOutputWithPoolingAndCrossAttentions

* Add cross attentions to TFBertModel

* Fix past or past_key_values argument issue

* Fix generation

* Fix save and load

* Add some checks and comments

* Clean the code that deals with past keys/values

* Add kwargs to processing_inputs

* Add serving_output to TFEncoderDecoderModel

* Some cleaning + fix use_cache value issue

* Fix tests + add bert2bert/bert2gpt2 tests

* Fix more tests

* Ignore crossattention.bias when loading GPT2 weights into TFGPT2

* Fix return_dict_in_generate in tf generation

* Fix is_token_logit_eos_token bug in tf generation

* Finalize the tests after fixing some bugs

* Fix another is_token_logit_eos_token bug in tf generation

* Add/Update docs

* Add TFBertEncoderDecoderModelTest

* Clean test script

* Add TFEncoderDecoderModel to the library

* Add cross attentions to TFRobertaModel

* Add TFRobertaEncoderDecoderModelTest

* make style

* Change the way of position_ids computation

* bug fix

* Fix copies in tf_albert

* Remove some copied from and apply some fix-copies

* Remove some copied

* Add cross attentions to some other TF models

* Remove encoder_hidden_states from TFLayoutLMModel.call for now

* Make style

* Fix TFRemBertForCausalLM

* Revert the change to longformer + Remove copies

* Revert the change to albert and convbert + Remove copies

* make quality

* make style

* Add TFRembertEncoderDecoderModelTest

* make quality and fix-copies

* test TFRobertaForCausalLM

* Fixes for failed tests

* Fixes for failed tests

* fix more tests

* Fixes for failed tests

* Fix Auto mapping order

* Fix TFRemBertEncoder return value

* fix tf_rembert

* Check copies are OK

* Fix missing TFBaseModelOutputWithPastAndCrossAttentions is not defined

* Add TFEncoderDecoderModelSaveLoadTests

* fix tf weight loading

* check the change of use_cache

* Revert the change

* Add missing test_for_causal_lm for TFRobertaModelTest

* Try cleaning past

* fix _reorder_cache

* Revert some files to original versions

* Keep as many copies as possible

* Apply suggested changes - Use raise ValueError instead of assert

* Move import to top

* Fix wrong require_torch

* Replace more assert by raise ValueError

* Add test_pt_tf_model_equivalence (the test won't pass for now)

* add test for loading/saving

* finish

* finish

* Remove test_pt_tf_model_equivalence

* Update tf modeling template

* Remove pooling, added in the prev. commit, from MainLayer

* Update tf modeling test template

* Move inputs["use_cache"] = False to modeling_tf_utils.py

* Fix torch.Tensor in the comment

* fix use_cache

* Fix missing use_cache in ElectraConfig

* Add a note to from_pretrained

* Fix style

* Change test_encoder_decoder_save_load_from_encoder_decoder_from_pt

* Fix TFMLP (in TFGPT2) activation issue

* Fix None past_key_values value in serving_output

* Don't call get_encoderdecoder_model in TFEncoderDecoderModelTest.test_configuration_tie until we have a TF checkpoint on Hub

* Apply review suggestions - style for cross_attns in serving_output

* Apply review suggestions - change assert + docstrings

* break the error message to respect the char limit

* deprecate the argument past

* fix docstring style

* Update the encoder-decoder rst file

* fix Unknown interpreted text role "method"

* fix typo

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

docs/source/model_doc/encoderdecoder.rst

@@ -27,6 +27,25 @@ An application of this architecture could be to leverage two pretrained :class:`
 and decoder for a summarization model as was shown in: `Text Summarization with Pretrained Encoders
 <https://arxiv.org/abs/1908.08345>`__ by Yang Liu and Mirella Lapata.
 
+The :meth:`~transformers.TFEncoderDecoderModel.from_pretrained` currently doesn't support initializing the model from a
+pytorch checkpoint. Passing ``from_pt=True`` to this method will throw an exception. If there are only pytorch
+checkpoints for a particular encoder-decoder model, a workaround is:
+
+.. code-block::
+
+    >>> # a workaround to load from pytorch checkpoint
+    >>> _model = EncoderDecoderModel.from_pretrained("patrickvonplaten/bert2bert-cnn_dailymail-fp16")
+    >>> _model.encoder.save_pretrained("./encoder")
+    >>> _model.decoder.save_pretrained("./decoder")
+    >>> model = TFEncoderDecoderModel.from_encoder_decoder_pretrained(
+    ...     "./encoder", "./decoder", encoder_from_pt=True, decoder_from_pt=True
+    ... )
+    >>> # This is only for copying some specific attributes of this particular model.
+    >>> model.config = _model.config
+
+This model was contributed by `thomwolf <https://github.com/thomwolf>`__. This model's TensorFlow and Flax versions
+were contributed by `ydshieh <https://github.com/ydshieh>`__.
+
 
 EncoderDecoderConfig
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
@@ -42,6 +61,13 @@ EncoderDecoderModel
     :members: forward, from_encoder_decoder_pretrained
 
 
+TFEncoderDecoderModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFEncoderDecoderModel
+    :members: call, from_encoder_decoder_pretrained
+
+
 FlaxEncoderDecoderModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

docs/source/model_doc/roberta.rst

@@ -126,6 +126,13 @@ TFRobertaModel
     :members: call
 
 
+TFRobertaForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFRobertaForCausalLM
+    :members: call
+
+
 TFRobertaForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~