TFMarian, TFMbart, TFPegasus, TFBlenderbot

docs/source/model_doc/blenderbot.rst

@@ -95,3 +95,12 @@ See :obj:`transformers.BartForConditionalGeneration` for arguments to `forward`
 
 .. autoclass:: transformers.BlenderbotForConditionalGeneration
     :members:
+
+
+TFBlenderbotForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+See :obj:`transformers.TFBartForConditionalGeneration` for arguments to `forward` and `generate`
+
+.. autoclass:: transformers.TFBlenderbotForConditionalGeneration
+    :members:

docs/source/model_doc/marian.rst

@@ -129,3 +129,9 @@ MarianMTModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.MarianMTModel
+
+
+TFMarianMTModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMarianMTModel

docs/source/model_doc/mbart.rst

@@ -79,4 +79,11 @@ MBartForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.MBartForConditionalGeneration
-    :members: forward
+    :members:
+
+
+TFMBartForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFMBartForConditionalGeneration
+    :members:

docs/source/model_doc/pegasus.rst

@@ -95,3 +95,9 @@ PegasusForConditionalGeneration
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.PegasusForConditionalGeneration
+
+
+TFPegasusForConditionalGeneration
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFPegasusForConditionalGeneration

src/transformers/__init__.py

@@ -670,6 +670,7 @@
         TFBertModel,
         TFBertPreTrainedModel,
     )
+    from .modeling_tf_blenderbot import TFBlenderbotForConditionalGeneration
     from .modeling_tf_camembert import (
         TF_CAMEMBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFCamembertForMaskedLM,
@@ -750,6 +751,8 @@
         TFLxmertPreTrainedModel,
         TFLxmertVisualFeatureEncoder,
     )
+    from .modeling_tf_marian import TFMarianMTModel
+    from .modeling_tf_mbart import TFMBartForConditionalGeneration
     from .modeling_tf_mobilebert import (
         TF_MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFMobileBertForMaskedLM,
@@ -771,6 +774,7 @@
         TFOpenAIGPTModel,
         TFOpenAIGPTPreTrainedModel,
     )
+    from .modeling_tf_pegasus import TFPegasusForConditionalGeneration
     from .modeling_tf_roberta import (
         TF_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFRobertaForMaskedLM,

src/transformers/modeling_bart.py

@@ -427,7 +427,6 @@ def forward(
         output_attentions=False,
     ):
         residual = x
-
         if layer_state is None:
             layer_state = {}
         if self.normalize_before:
@@ -447,7 +446,7 @@ def forward(
         if not self.normalize_before:
             x = self.self_attn_layer_norm(x)
 
-        # Cross attention
+        # Cross-Attention Block
         residual = x
         assert self.encoder_attn.cache_key != self.self_attn.cache_key
         if self.normalize_before:
@@ -628,7 +627,6 @@ def forward(
         encoder_hidden_states = encoder_hidden_states.transpose(0, 1)
 
         next_cache = next_decoder_cache if use_cache else None
-
         if not return_dict:
             return tuple(v for v in [x, next_cache, all_hidden_states, all_self_attns] if v is not None)
         return BaseModelOutputWithPast(

src/transformers/modeling_tf_auto.py

@@ -41,6 +41,10 @@
     XLNetConfig,
     replace_list_option_in_docstrings,
 )
+from .configuration_blenderbot import BlenderbotConfig
+from .configuration_marian import MarianConfig
+from .configuration_mbart import MBartConfig
+from .configuration_pegasus import PegasusConfig
 from .configuration_utils import PretrainedConfig
 from .file_utils import add_start_docstrings
 from .modeling_tf_albert import (
@@ -63,6 +67,7 @@
     TFBertLMHeadModel,
     TFBertModel,
 )
+from .modeling_tf_blenderbot import TFBlenderbotForConditionalGeneration
 from .modeling_tf_camembert import (
     TFCamembertForMaskedLM,
     TFCamembertForMultipleChoice,
@@ -108,6 +113,8 @@
 )
 from .modeling_tf_gpt2 import TFGPT2LMHeadModel, TFGPT2Model
 from .modeling_tf_longformer import TFLongformerForMaskedLM, TFLongformerForQuestionAnswering, TFLongformerModel
+from .modeling_tf_marian import TFMarianMTModel
+from .modeling_tf_mbart import TFMBartForConditionalGeneration
 from .modeling_tf_mobilebert import (
     TFMobileBertForMaskedLM,
     TFMobileBertForMultipleChoice,
@@ -118,6 +125,7 @@
     TFMobileBertModel,
 )
 from .modeling_tf_openai import TFOpenAIGPTLMHeadModel, TFOpenAIGPTModel
+from .modeling_tf_pegasus import TFPegasusForConditionalGeneration
 from .modeling_tf_roberta import (
     TFRobertaForMaskedLM,
     TFRobertaForMultipleChoice,
@@ -210,6 +218,7 @@
         (T5Config, TFT5ForConditionalGeneration),
         (DistilBertConfig, TFDistilBertForMaskedLM),
         (AlbertConfig, TFAlbertForMaskedLM),
+        (MarianConfig, TFMarianMTModel),
         (BartConfig, TFBartForConditionalGeneration),
         (CamembertConfig, TFCamembertForMaskedLM),
         (XLMRobertaConfig, TFXLMRobertaForMaskedLM),
@@ -261,8 +270,16 @@
     ]
 )
 
+
 TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING = OrderedDict(
-    [(T5Config, TFT5ForConditionalGeneration), (BartConfig, TFBartForConditionalGeneration)]
+    [
+        (T5Config, TFT5ForConditionalGeneration),
+        (MarianConfig, TFMarianMTModel),
+        (MBartConfig, TFMBartForConditionalGeneration),
+        (PegasusConfig, TFPegasusForConditionalGeneration),
+        (BlenderbotConfig, TFBlenderbotForConditionalGeneration),
+        (BartConfig, TFBartForConditionalGeneration),
+    ]
 )
 
 TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(