[All models] fix docs after adding output attentions to all forward functions (#4909)

* fix doc

* add format file

* add output attentions to all docs

* add also for bart

* fix naming

* re-add doc to config

src/transformers/configuration_utils.py

@@ -44,8 +44,10 @@ class PretrainedConfig(object):
                 Name of the task used to fine-tune the model. This can be used when converting from an original (TensorFlow or PyTorch) checkpoint.
             num_labels (:obj:`int`, `optional`, defaults to `2`):
                 Number of classes to use when the model is a classification model (sequences/tokens)
-            output_hidden_states (:obj:`string`, `optional`, defaults to :obj:`False`):
+            output_hidden_states (:obj:`bool`, `optional`, defaults to :obj:`False`):
                 Should the model returns all hidden-states.
+            output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Should the model returns all attentions.
             torchscript (:obj:`bool`, `optional`, defaults to :obj:`False`):
                 Is the model used with Torchscript (for PyTorch models).
     """

src/transformers/modeling_albert.py

@@ -423,6 +423,8 @@ ALBERT_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -507,7 +509,7 @@ class AlbertModel(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -631,7 +633,7 @@ class AlbertForPreTraining(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -778,7 +780,7 @@ class AlbertForMaskedLM(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -873,7 +875,7 @@ class AlbertForSequenceClassification(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -968,7 +970,7 @@ class AlbertForTokenClassification(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1074,7 +1076,7 @@ class AlbertForQuestionAnswering(AlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_bart.py

@@ -89,6 +89,8 @@ BART_INPUTS_DOCSTRING = r"""
             Default behavior: generate a tensor that ignores pad tokens in decoder_input_ids. Causal mask will also be used by default.
             If you want to change padding behavior, you should read :func:`~transformers.modeling_bart._prepare_decoder_inputs` and modify.
             See diagram 1 in the paper for more info on the default strategy
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -928,7 +930,7 @@ class BartForConditionalGeneration(PretrainedBartModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1078,7 +1080,7 @@ class BartForSequenceClassification(PretrainedBartModel):
                 Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
                 of shape :obj:`(batch_size, sequence_length, hidden_size)`.
                 Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-            attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+            attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
                 Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
                 Attentions weights after the attention softmax, used to compute the weighted average in the
                 self-attention

src/transformers/modeling_bert.py

@@ -594,6 +594,8 @@ BERT_INPUTS_DOCSTRING = r"""
             is used in the cross-attention if the model is configured as a decoder.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -674,7 +676,7 @@ class BertModel(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -814,7 +816,7 @@ class BertForPreTraining(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -928,7 +930,7 @@ class BertForMaskedLM(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1060,7 +1062,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1152,7 +1154,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1251,7 +1253,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1354,7 +1356,7 @@ class BertForTokenClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1461,7 +1463,7 @@ class BertForQuestionAnswering(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_camembert.py

@@ -49,6 +49,8 @@ CAMEMBERT_START_DOCSTRING = r"""
             model. Initializing with a config file does not load the weights associated with the model, only the
             configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 

src/transformers/modeling_ctrl.py

@@ -266,6 +266,8 @@ CTRL_INPUTS_DOCSTRING = r"""
         use_cache (:obj:`bool`):
             If `use_cache` is True, `past` key value states are returned and
             can be used to speed up decoding (see `past`). Defaults to `True`.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -332,7 +334,7 @@ class CTRLModel(CTRLPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -521,7 +523,7 @@ class CTRLLMHeadModel(CTRLPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_distilbert.py

@@ -377,6 +377,8 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -421,7 +423,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -523,7 +525,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -616,7 +618,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -713,7 +715,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -812,7 +814,7 @@ class DistilBertForTokenClassification(DistilBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_electra.py

@@ -220,6 +220,8 @@ ELECTRA_INPUTS_DOCSTRING = r"""
             is used in the cross-attention if the model is configured as a decoder.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -281,7 +283,7 @@ class ElectraModel(ElectraPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -404,7 +406,7 @@ class ElectraForSequenceClassification(ElectraPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -494,7 +496,7 @@ class ElectraForPreTraining(ElectraPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -597,7 +599,7 @@ class ElectraForMaskedLM(ElectraPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -691,7 +693,7 @@ class ElectraForTokenClassification(ElectraPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_flaubert.py

@@ -100,6 +100,8 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -140,7 +142,7 @@ class FlaubertModel(XLMModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_gpt2.py

@@ -335,6 +335,8 @@ GPT2_INPUTS_DOCSTRING = r"""
             If `past` is used, optionally only the last `inputs_embeds` have to be input (see `past`).
         use_cache (:obj:`bool`):
             If `use_cache` is True, `past` key value states are returned and can be used to speed up decoding (see `past`). Defaults to `True`.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -395,7 +397,7 @@ class GPT2Model(GPT2PreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -582,7 +584,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -703,7 +705,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_longformer.py

@@ -487,6 +487,8 @@ LONGFORMER_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -597,7 +599,7 @@ class LongformerModel(RobertaModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -726,7 +728,7 @@ class LongformerForMaskedLM(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -825,7 +827,7 @@ class LongformerForSequenceClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -948,7 +950,7 @@ class LongformerForQuestionAnswering(BertPreTrainedModel):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
@@ -1069,7 +1071,7 @@ class LongformerForTokenClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1172,7 +1174,7 @@ class LongformerForMultipleChoice(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_mmbt.py

@@ -141,6 +141,8 @@ MMBT_INPUTS_DOCSTRING = r"""    Inputs:
             is used in the cross-attention if the model is configured as a decoder.
             Mask values selected in ``[0, 1]``:
             ``1`` for tokens that are NOT MASKED, ``0`` for MASKED tokens.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 

src/transformers/modeling_openai.py

@@ -324,6 +324,8 @@ OPENAI_GPT_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Should the model returns attentions weights.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -377,7 +379,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -521,7 +523,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -638,7 +640,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_reformer.py

@@ -1505,6 +1505,8 @@ REFORMER_INPUTS_DOCSTRING = r"""
             bucketing. Setting `num_hashes` overwrites the default `num_hashes` defined
             in `config.num_hashes`.
             For more information, see `num_hashes` in :class:`transformers.ReformerConfig`.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -1561,7 +1563,7 @@ class ReformerModel(ReformerPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        all_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        all_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1763,7 +1765,7 @@ class ReformerModelWithLMHead(ReformerPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        all_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        all_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_roberta.py

@@ -130,6 +130,8 @@ ROBERTA_INPUTS_DOCSTRING = r"""
             Optionally, instead of passing :obj:`input_ids` you can choose to directly pass an embedded representation.
             This is useful if you want more control over how to convert `input_ids` indices into associated vectors
             than the model's internal embedding lookup matrix.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -208,7 +210,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -328,7 +330,7 @@ class RobertaForSequenceClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -423,7 +425,7 @@ class RobertaForMultipleChoice(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -521,7 +523,7 @@ class RobertaForTokenClassification(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -649,7 +651,7 @@ class RobertaForQuestionAnswering(BertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_t5.py

@@ -841,6 +841,8 @@ T5_INPUTS_DOCSTRING = r"""
             Mask to nullify selected heads of the self-attention modules.
             Mask values selected in ``[0, 1]``:
             ``1`` indicates the head is **not masked**, ``0`` indicates the head is **masked**.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -914,7 +916,7 @@ class T5Model(T5PreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1050,7 +1052,7 @@ class T5ForConditionalGeneration(T5PreTrainedModel):
             Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention.

src/transformers/modeling_tf_albert.py

@@ -688,6 +688,8 @@ ALBERT_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -719,7 +721,7 @@ class TFAlbertModel(TFAlbertPreTrainedModel):
                 of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
                 Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-            attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+            attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
                 tuple of :obj:`tf.Tensor` (one for each layer) of shape
                 :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -771,7 +773,7 @@ class TFAlbertForPreTraining(TFAlbertPreTrainedModel):
             tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
             Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads.
@@ -831,7 +833,7 @@ class TFAlbertForMaskedLM(TFAlbertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -905,7 +907,7 @@ class TFAlbertForSequenceClassification(TFAlbertPreTrainedModel, TFSequenceClass
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -993,7 +995,7 @@ class TFAlbertForTokenClassification(TFAlbertPreTrainedModel, TFTokenClassificat
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -1089,7 +1091,7 @@ class TFAlbertForQuestionAnswering(TFAlbertPreTrainedModel, TFQuestionAnsweringL
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -1196,7 +1198,7 @@ class TFAlbertForMultipleChoice(TFAlbertPreTrainedModel, TFMultipleChoiceLoss):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 

src/transformers/modeling_tf_bert.py

@@ -682,6 +682,8 @@ BERT_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -713,7 +715,7 @@ class TFBertModel(TFBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -765,7 +767,7 @@ class TFBertForPreTraining(TFBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -819,7 +821,7 @@ class TFBertForMaskedLM(TFBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -869,7 +871,7 @@ class TFBertForNextSentencePrediction(TFBertPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -945,7 +947,7 @@ class TFBertForSequenceClassification(TFBertPreTrainedModel, TFSequenceClassific
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -1044,7 +1046,7 @@ class TFBertForMultipleChoice(TFBertPreTrainedModel, TFMultipleChoiceLoss):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -1169,7 +1171,7 @@ class TFBertForTokenClassification(TFBertPreTrainedModel, TFTokenClassificationL
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 
@@ -1266,7 +1268,7 @@ class TFBertForQuestionAnswering(TFBertPreTrainedModel, TFQuestionAnsweringLoss)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`:
 

src/transformers/modeling_tf_camembert.py

@@ -62,6 +62,8 @@ CAMEMBERT_START_DOCSTRING = r"""
         config (:class:`~transformers.CamembertConfig`): Model configuration class with all the parameters of the
             model. Initializing with a config file does not load the weights associated with the model, only the configuration.
             Check out the :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model weights.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 

src/transformers/modeling_tf_ctrl.py

@@ -464,6 +464,8 @@ CTRL_INPUTS_DOCSTRING = r"""
         training (:obj:`boolean`, `optional`, defaults to :obj:`False`):
             Whether to activate dropout modules (if set to :obj:`True`) during training or to de-activate them
             (if set to :obj:`False`) for evaluation.
+        output_attentions (:obj:`bool`, `optional`, defaults to `:obj:`None`):
+            If set to ``True``, the attentions tensors of all attention layers are returned. See ``attentions`` under returned tensors for more detail.
 """
 
 
@@ -492,7 +494,7 @@ class TFCTRLModel(TFCTRLPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -572,7 +574,7 @@ class TFCTRLLMHeadModel(TFCTRLPreTrainedModel):
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.
 
             Hidden-states of the model at the output of each layer plus the initial embedding outputs.
-        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True``):
+        attentions (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``output_attentions=True`` is passed or ``config.output_attentions=True``):
             Tuple of :obj:`tf.Tensor` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.