[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/modeling_tf_distilbert.py

@@ -549,6 +549,8 @@ DISTILBERT_INPUTS_DOCSTRING = r"""
             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.
 """
 
 
@@ -573,7 +575,7 @@ class TFDistilBertModel(TFDistilBertPreTrainedModel):
             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)`:
 
@@ -646,7 +648,7 @@ class TFDistilBertForMaskedLM(TFDistilBertPreTrainedModel):
             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)`:
 
@@ -725,7 +727,7 @@ class TFDistilBertForSequenceClassification(TFDistilBertPreTrainedModel, TFSeque
             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)`:
 
@@ -809,7 +811,7 @@ class TFDistilBertForTokenClassification(TFDistilBertPreTrainedModel, TFTokenCla
             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)`:
 
@@ -909,7 +911,7 @@ class TFDistilBertForMultipleChoice(TFDistilBertPreTrainedModel, TFMultipleChoic
             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)`:
 
@@ -1032,7 +1034,7 @@ class TFDistilBertForQuestionAnswering(TFDistilBertPreTrainedModel, TFQuestionAn
             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_electra.py

@@ -348,6 +348,8 @@ ELECTRA_INPUTS_DOCSTRING = r"""
             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.
 """
 
 
@@ -379,7 +381,7 @@ class TFElectraModel(TFElectraPreTrainedModel):
             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)`:
 
@@ -441,7 +443,7 @@ class TFElectraForPreTraining(TFElectraPreTrainedModel):
             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)`:
 
@@ -542,7 +544,7 @@ class TFElectraForMaskedLM(TFElectraPreTrainedModel):
             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)`:
 
@@ -624,7 +626,7 @@ class TFElectraForTokenClassification(TFElectraPreTrainedModel, TFTokenClassific
             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)`:
 
@@ -720,7 +722,7 @@ class TFElectraForQuestionAnswering(TFElectraPreTrainedModel, TFQuestionAnswerin
             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_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.
 """
 
 

src/transformers/modeling_tf_gpt2.py

@@ -467,6 +467,8 @@ GPT2_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.
 """
 
 
@@ -495,7 +497,7 @@ class TFGPT2Model(TFGPT2PreTrainedModel):
             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)`.
 
@@ -554,7 +556,7 @@ class TFGPT2LMHeadModel(TFGPT2PreTrainedModel):
             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)`.
 
@@ -639,7 +641,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
             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_openai.py

@@ -429,6 +429,8 @@ OPENAI_GPT_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.
 """
 
 
@@ -453,7 +455,7 @@ class TFOpenAIGPTModel(TFOpenAIGPTPreTrainedModel):
             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)`.
 
@@ -501,7 +503,7 @@ class TFOpenAIGPTLMHeadModel(TFOpenAIGPTPreTrainedModel):
             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)`.
 
@@ -583,7 +585,7 @@ class TFOpenAIGPTDoubleHeadsModel(TFOpenAIGPTPreTrainedModel):
             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_roberta.py

@@ -182,6 +182,8 @@ ROBERTA_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.
 """
 
 
@@ -213,7 +215,7 @@ class TFRobertaModel(TFRobertaPreTrainedModel):
             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)`:
 
@@ -289,7 +291,7 @@ class TFRobertaForMaskedLM(TFRobertaPreTrainedModel):
             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)`:
 
@@ -378,7 +380,7 @@ class TFRobertaForSequenceClassification(TFRobertaPreTrainedModel, TFSequenceCla
             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)`:
 
@@ -474,7 +476,7 @@ class TFRobertaForMultipleChoice(TFRobertaPreTrainedModel, 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)`:
 
@@ -596,7 +598,7 @@ class TFRobertaForTokenClassification(TFRobertaPreTrainedModel, TFTokenClassific
             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)`:
 
@@ -692,7 +694,7 @@ class TFRobertaForQuestionAnswering(TFRobertaPreTrainedModel, TFQuestionAnswerin
             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_t5.py

@@ -855,6 +855,8 @@ 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.
 """
 
 
@@ -909,7 +911,7 @@ class TFT5Model(TFT5PreTrainedModel):
             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)`.
 
@@ -1037,7 +1039,7 @@ class TFT5ForConditionalGeneration(TFT5PreTrainedModel):
             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_transfo_xl.py

@@ -692,6 +692,8 @@ TRANSFO_XL_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.
 """
 
 
@@ -720,7 +722,7 @@ class TFTransfoXLModel(TFTransfoXLPreTrainedModel):
             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)`.
 
@@ -818,7 +820,7 @@ class TFTransfoXLLMHeadModel(TFTransfoXLPreTrainedModel):
             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_xlm.py

@@ -585,6 +585,8 @@ XLM_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.
 """
 
 
@@ -609,7 +611,7 @@ class TFXLMModel(TFXLMPreTrainedModel):
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -705,7 +707,7 @@ class TFXLMWithLMHeadModel(TFXLMPreTrainedModel):
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -778,7 +780,7 @@ class TFXLMForSequenceClassification(TFXLMPreTrainedModel, TFSequenceClassificat
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -878,7 +880,7 @@ class TFXLMForMultipleChoice(TFXLMPreTrainedModel, 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)`:
 
@@ -1012,7 +1014,7 @@ class TFXLMForTokenClassification(TFXLMPreTrainedModel, TFTokenClassificationLos
             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)`:
 
@@ -1110,7 +1112,7 @@ class TFXLMForQuestionAnsweringSimple(TFXLMPreTrainedModel, 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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_tf_xlm_roberta.py

@@ -62,6 +62,8 @@ XLM_ROBERTA_START_DOCSTRING = r"""
         config (:class:`~transformers.XLMRobertaConfig`): 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_xlnet.py

@@ -809,6 +809,8 @@ XLNET_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         use_cache (:obj:`bool`):
             If `use_cache` is True, `mems` are returned and can be used to speed up decoding (see `mems`). 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.
 """
 
 
@@ -837,7 +839,7 @@ class TFXLNetModel(TFXLNetPreTrainedModel):
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -921,7 +923,7 @@ class TFXLNetLMHeadModel(TFXLNetPreTrainedModel):
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1012,7 +1014,7 @@ class TFXLNetForSequenceClassification(TFXLNetPreTrainedModel, TFSequenceClassif
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1119,7 +1121,7 @@ class TFXLNetForMultipleChoice(TFXLNetPreTrainedModel, 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)`:
 
@@ -1260,7 +1262,7 @@ class TFXLNetForTokenClassification(TFXLNetPreTrainedModel, TFTokenClassificatio
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 
@@ -1368,7 +1370,7 @@ class TFXLNetForQuestionAnsweringSimple(TFXLNetPreTrainedModel, TFQuestionAnswer
             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` or :obj:`Numpy array` (one for each layer) of shape
             :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
 

src/transformers/modeling_transfo_xl.py

@@ -542,6 +542,8 @@ TRANSFO_XL_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.
 """
 
 
@@ -683,7 +685,7 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
             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)`.
 
@@ -879,7 +881,7 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
             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_xlm.py

@@ -302,6 +302,8 @@ XLM_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.
 """
 
 
@@ -417,7 +419,7 @@ class XLMModel(XLMPreTrainedModel):
             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)`.
 
@@ -660,7 +662,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
             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)`.
 
@@ -747,7 +749,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
             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)`.
 
@@ -851,7 +853,7 @@ class XLMForQuestionAnsweringSimple(XLMPreTrainedModel):
             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)`.
 
@@ -987,7 +989,7 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
             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)`.
 
@@ -1081,7 +1083,7 @@ class XLMForTokenClassification(XLMPreTrainedModel):
             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_xlm_roberta.py

@@ -53,6 +53,8 @@ XLM_ROBERTA_START_DOCSTRING = r"""
         config (:class:`~transformers.XLMRobertaConfig`): 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_xlnet.py

@@ -618,6 +618,8 @@ XLNET_INPUTS_DOCSTRING = r"""
             than the model's internal embedding lookup matrix.
         use_cache (:obj:`bool`):
             If `use_cache` is True, `mems` are returned and can be used to speed up decoding (see `mems`). 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.
 """
 
 
@@ -777,7 +779,7 @@ class XLNetModel(XLNetPreTrainedModel):
             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)`.
 
@@ -1075,7 +1077,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
             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)`.
 
@@ -1194,7 +1196,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
             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)`.
 
@@ -1300,7 +1302,7 @@ class XLNetForTokenClassification(XLNetPreTrainedModel):
             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)`.
 
@@ -1413,7 +1415,7 @@ class XLNetForMultipleChoice(XLNetPreTrainedModel):
             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)`.
 
@@ -1533,7 +1535,7 @@ class XLNetForQuestionAnsweringSimple(XLNetPreTrainedModel):
             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)`.
 
@@ -1677,7 +1679,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
             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)`.