Add caching mechanism to BERT, RoBERTa (#9183)

* add past_key_values

* add use_cache option

* make mask before cutting ids

* adjust position_ids according to past_key_values

* flatten past_key_values

* fix positional embeds

* fix _reorder_cache

* set use_cache to false when not decoder, fix attention mask init

* add test for caching

* add past_key_values for Roberta

* fix position embeds

* add caching test for roberta

* add doc

* make style

* doc, fix attention mask, test

* small fixes

* adress patrick's comments

* input_ids shouldn't start with pad token

* use_cache only when decoder

* make consistent with bert

* make copies consistent

* add use_cache to encoder

* add past_key_values to tapas attention

* apply suggestions from code review

* make coppies consistent

* add attn mask in tests

* remove copied from longformer

* apply suggestions from code review

* fix bart test

* nit

* simplify model outputs

* fix doc

* fix output ordering

docs/source/main_classes/output.rst

@@ -126,13 +126,6 @@ CausalLMOutputWithCrossAttentions
     :members:
 
 
-CausalLMOutputWithPastAndCrossAttentions
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-.. autoclass:: transformers.modeling_outputs.CausalLMOutputWithPastAndCrossAttentions
-    :members:
-
-
 CausalLMOutputWithPast
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 

src/transformers/modeling_outputs.py

@@ -175,11 +175,19 @@ class BaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):
 
             Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
             weighted average in the cross-attention heads.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`torch.FloatTensor` tuples of length :obj:`config.n_layers`, with each tuple containing the
+            cached key, value states of the self-attention and the cross-attention layers if model is used in
+            encoder-decoder setting. Only relevant if ``config.is_decoder = True``.
+
+            Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
+            :obj:`past_key_values` input) to speed up sequential decoding.
     """
 
     last_hidden_state: torch.FloatTensor = None
     pooler_output: torch.FloatTensor = None
     hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
     attentions: Optional[Tuple[torch.FloatTensor]] = None
     cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
 
@@ -379,53 +387,18 @@ class CausalLMOutputWithCrossAttentions(ModelOutput):
 
             Cross attentions weights after the attention softmax, used to compute the weighted average in the
             cross-attention heads.
-    """
-
-    loss: Optional[torch.FloatTensor] = None
-    logits: torch.FloatTensor = None
-    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
-    attentions: Optional[Tuple[torch.FloatTensor]] = None
-    cross_attentions: Optional[Tuple[torch.FloatTensor]] = None
-
-
-@dataclass
-class CausalLMOutputWithPastAndCrossAttentions(ModelOutput):
-    """
-    Base class for causal language model (or autoregressive) outputs.
-
-    Args:
-        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`labels` is provided):
-            Language modeling loss (for next-token prediction).
-        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, config.vocab_size)`):
-            Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
-        past_key_values (:obj:`List[torch.FloatTensor]`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
-            List of :obj:`torch.FloatTensor` of length :obj:`config.n_layers`, with each tensor of shape :obj:`(2,
-            batch_size, num_heads, sequence_length, embed_size_per_head)`).
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))`, `optional`, returned when ``use_cache=True`` is passed or when ``config.use_cache=True``):
+            Tuple of :obj:`torch.FloatTensor` tuples of length :obj:`config.n_layers`, with each tuple containing the
+            cached key, value states of the self-attention and the cross-attention layers if model is used in
+            encoder-decoder setting. Only relevant if ``config.is_decoder = True``.
 
             Contains pre-computed hidden-states (key and values in the attention blocks) that can be used (see
             :obj:`past_key_values` input) to speed up sequential decoding.
-        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
-            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`` is passed or when ``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
-            heads.
-        cross_attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
-            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape :obj:`(batch_size, num_heads,
-            sequence_length, sequence_length)`.
-
-            Cross attentions weights after the attention softmax, used to compute the weighted average in the
-            cross-attention heads.
     """
 
     loss: Optional[torch.FloatTensor] = None
     logits: torch.FloatTensor = None
-    past_key_values: Optional[List[torch.FloatTensor]] = None
+    past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
     hidden_states: Optional[Tuple[torch.FloatTensor]] = None
     attentions: Optional[Tuple[torch.FloatTensor]] = None
     cross_attentions: Optional[Tuple[torch.FloatTensor]] = None

src/transformers/models/albert/modeling_albert.py

@@ -217,7 +217,9 @@ class AlbertEmbeddings(nn.Module):
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
 
     # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
-    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
         if input_ids is not None:
             input_shape = input_ids.size()
         else:
@@ -226,7 +228,7 @@ class AlbertEmbeddings(nn.Module):
         seq_length = input_shape[1]
 
         if position_ids is None:
-            position_ids = self.position_ids[:, :seq_length]
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
 
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)

src/transformers/models/bert/configuration_bert.py

@@ -98,6 +98,9 @@ class BertConfig(PretrainedConfig):
             <https://arxiv.org/abs/1803.02155>`__. For more information on :obj:`"relative_key_query"`, please refer to
             `Method 4` in `Improve Transformer Models with Better Relative Position Embeddings (Huang et al.)
             <https://arxiv.org/abs/2009.13658>`__.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if ``config.is_decoder=True``.
 
     Examples::
 
@@ -131,6 +134,7 @@ class BertConfig(PretrainedConfig):
         pad_token_id=0,
         gradient_checkpointing=False,
         position_embedding_type="absolute",
+        use_cache=True,
         **kwargs
     ):
         super().__init__(pad_token_id=pad_token_id, **kwargs)
@@ -149,3 +153,4 @@ class BertConfig(PretrainedConfig):
         self.layer_norm_eps = layer_norm_eps
         self.gradient_checkpointing = gradient_checkpointing
         self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache

src/transformers/models/bert/modeling_bert.py

@@ -36,7 +36,7 @@ from ...file_utils import (
     replace_return_docstrings,
 )
 from ...modeling_outputs import (
-    BaseModelOutputWithCrossAttentions,
+    BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
     CausalLMOutputWithCrossAttentions,
     MaskedLMOutput,
@@ -180,7 +180,9 @@ class BertEmbeddings(nn.Module):
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
 
-    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
         if input_ids is not None:
             input_shape = input_ids.size()
         else:
@@ -189,7 +191,7 @@ class BertEmbeddings(nn.Module):
         seq_length = input_shape[1]
 
         if position_ids is None:
-            position_ids = self.position_ids[:, :seq_length]
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
 
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
@@ -230,6 +232,8 @@ class BertSelfAttention(nn.Module):
             self.max_position_embeddings = config.max_position_embeddings
             self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
 
+        self.is_decoder = config.is_decoder
+
     def transpose_for_scores(self, x):
         new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
         x = x.view(*new_x_shape)
@@ -242,6 +246,7 @@ class BertSelfAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         mixed_query_layer = self.query(hidden_states)
@@ -249,17 +254,37 @@ class BertSelfAttention(nn.Module):
         # If this is instantiated as a cross-attention module, the keys
         # and values come from an encoder; the attention mask needs to be
         # such that the encoder's padding tokens are not attended to.
-        if encoder_hidden_states is not None:
-            mixed_key_layer = self.key(encoder_hidden_states)
-            mixed_value_layer = self.value(encoder_hidden_states)
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
             attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
         else:
-            mixed_key_layer = self.key(hidden_states)
-            mixed_value_layer = self.value(hidden_states)
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
-        key_layer = self.transpose_for_scores(mixed_key_layer)
-        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
@@ -303,6 +328,9 @@ class BertSelfAttention(nn.Module):
         context_layer = context_layer.view(*new_context_layer_shape)
 
         outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
         return outputs
 
 
@@ -352,6 +380,7 @@ class BertAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         self_outputs = self.self(
@@ -360,6 +389,7 @@ class BertAttention(nn.Module):
             head_mask,
             encoder_hidden_states,
             encoder_attention_mask,
+            past_key_value,
             output_attentions,
         )
         attention_output = self.output(self_outputs[0], hidden_states)
@@ -417,36 +447,60 @@ class BertLayer(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
         self_attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             head_mask,
             output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
         )
         attention_output = self_attention_outputs[0]
-        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
 
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
         if self.is_decoder and encoder_hidden_states is not None:
             assert hasattr(
                 self, "crossattention"
             ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
             cross_attention_outputs = self.crossattention(
                 attention_output,
                 attention_mask,
                 head_mask,
                 encoder_hidden_states,
                 encoder_attention_mask,
+                cross_attn_past_key_value,
                 output_attentions,
             )
             attention_output = cross_attention_outputs[0]
-            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
 
         layer_output = apply_chunking_to_forward(
             self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
         )
         outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
         return outputs
 
     def feed_forward_chunk(self, attention_output):
@@ -468,6 +522,8 @@ class BertEncoder(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=False,
         output_hidden_states=False,
         return_dict=True,
@@ -475,17 +531,19 @@ class BertEncoder(nn.Module):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
-
+            past_key_value = past_key_values[i] if past_key_values is not None else None
             if getattr(self.config, "gradient_checkpointing", False):
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
+                        return module(*inputs, past_key_value, output_attentions)
 
                     return custom_forward
 
@@ -504,9 +562,13 @@ class BertEncoder(nn.Module):
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
                     output_attentions,
                 )
+
             hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
             if output_attentions:
                 all_self_attentions = all_self_attentions + (layer_outputs[1],)
                 if self.config.add_cross_attention:
@@ -518,11 +580,18 @@ class BertEncoder(nn.Module):
         if not return_dict:
             return tuple(
                 v
-                for v in [hidden_states, all_hidden_states, all_self_attentions, all_cross_attentions]
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
                 if v is not None
             )
-        return BaseModelOutputWithCrossAttentions(
+        return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
             hidden_states=all_hidden_states,
             attentions=all_self_attentions,
             cross_attentions=all_cross_attentions,
@@ -799,6 +868,8 @@ class BertModel(BertPreTrainedModel):
         inputs_embeds=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
         return_dict=None,
@@ -813,6 +884,15 @@ class BertModel(BertPreTrainedModel):
 
             - 1 for tokens that are **not masked**,
             - 0 for tokens that are **masked**.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -820,19 +900,29 @@ class BertModel(BertPreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
         if input_ids is not None and inputs_embeds is not None:
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
             input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         device = input_ids.device if input_ids is not None else inputs_embeds.device
 
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
         if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
 
@@ -859,7 +949,11 @@ class BertModel(BertPreTrainedModel):
         head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
 
         embedding_output = self.embeddings(
-            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
         )
         encoder_outputs = self.encoder(
             embedding_output,
@@ -867,6 +961,8 @@ class BertModel(BertPreTrainedModel):
             head_mask=head_mask,
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -880,6 +976,7 @@ class BertModel(BertPreTrainedModel):
         return BaseModelOutputWithPoolingAndCrossAttentions(
             last_hidden_state=sequence_output,
             pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
             hidden_states=encoder_outputs.hidden_states,
             attentions=encoder_outputs.attentions,
             cross_attentions=encoder_outputs.cross_attentions,
@@ -1029,6 +1126,8 @@ class BertLMHeadModel(BertPreTrainedModel):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         labels=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
         return_dict=None,
@@ -1047,6 +1146,15 @@ class BertLMHeadModel(BertPreTrainedModel):
             Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
             ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
             ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
 
         Returns:
 
@@ -1066,6 +1174,8 @@ class BertLMHeadModel(BertPreTrainedModel):
             >>> prediction_logits = outputs.logits
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
 
         outputs = self.bert(
             input_ids,
@@ -1076,6 +1186,8 @@ class BertLMHeadModel(BertPreTrainedModel):
             inputs_embeds=inputs_embeds,
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -1099,20 +1211,30 @@ class BertLMHeadModel(BertPreTrainedModel):
         return CausalLMOutputWithCrossAttentions(
             loss=lm_loss,
             logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
             cross_attentions=outputs.cross_attentions,
         )
 
-    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
         input_shape = input_ids.shape
-
         # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
         if attention_mask is None:
             attention_mask = input_ids.new_ones(input_shape)
 
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
         return {"input_ids": input_ids, "attention_mask": attention_mask}
 
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
 
 @add_start_docstrings("""Bert Model with a `language modeling` head on top. """, BERT_START_DOCSTRING)
 class BertForMaskedLM(BertPreTrainedModel):

src/transformers/models/bert_generation/modeling_bert_generation.py

@@ -26,7 +26,7 @@ from ...file_utils import (
     add_start_docstrings_to_model_forward,
     replace_return_docstrings,
 )
-from ...modeling_outputs import BaseModelOutputWithCrossAttentions, CausalLMOutputWithCrossAttentions
+from ...modeling_outputs import BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
 from ...modeling_utils import PreTrainedModel
 from ...utils import logging
 from ..bert.modeling_bert import BertEncoder
@@ -144,7 +144,7 @@ class BertGenerationEmbeddings(nn.Module):
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
         self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
 
-    def forward(self, input_ids=None, position_ids=None, inputs_embeds=None):
+    def forward(self, input_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0):
         if input_ids is not None:
             input_shape = input_ids.size()
         else:
@@ -153,7 +153,7 @@ class BertGenerationEmbeddings(nn.Module):
         seq_length = input_shape[1]
 
         if position_ids is None:
-            position_ids = self.position_ids[:, :seq_length]
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
 
         if inputs_embeds is None:
             inputs_embeds = self.word_embeddings(input_ids)
@@ -297,7 +297,7 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
         checkpoint="google/bert_for_seq_generation_L-24_bbc_encoder",
-        output_type=BaseModelOutputWithCrossAttentions,
+        output_type=BaseModelOutputWithPastAndCrossAttentions,
         config_class=_CONFIG_FOR_DOC,
     )
     def forward(
@@ -309,6 +309,8 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
         inputs_embeds=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
         return_dict=None,
@@ -321,6 +323,15 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
             Mask to avoid performing attention on the padding token indices of the encoder input. This mask 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.
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
         """
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -332,19 +343,28 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
             raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
         elif input_ids is not None:
             input_shape = input_ids.size()
+            batch_size, seq_length = input_shape
         elif inputs_embeds is not None:
             input_shape = inputs_embeds.size()[:-1]
+            batch_size, seq_length = input_shape
         else:
             raise ValueError("You have to specify either input_ids or inputs_embeds")
 
         device = input_ids.device if input_ids is not None else inputs_embeds.device
 
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
         if attention_mask is None:
-            attention_mask = torch.ones(input_shape, device=device)
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
 
         # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
         # ourselves in which case we just need to make it broadcastable to all heads.
-        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+        extended_attention_mask = None
+        if not use_cache:
+            extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(
+                attention_mask, input_shape, device
+            )
 
         # If a 2D or 3D attention mask is provided for the cross-attention
         # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
@@ -364,7 +384,12 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
         # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
         head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
 
-        embedding_output = self.embeddings(input_ids=input_ids, position_ids=position_ids, inputs_embeds=inputs_embeds)
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
 
         encoder_outputs = self.encoder(
             embedding_output,
@@ -372,6 +397,8 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
             head_mask=head_mask,
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -381,8 +408,9 @@ class BertGenerationEncoder(BertGenerationPreTrainedModel):
         if not return_dict:
             return (sequence_output,) + encoder_outputs[1:]
 
-        return BaseModelOutputWithCrossAttentions(
+        return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=sequence_output,
+            past_key_values=encoder_outputs.past_key_values,
             hidden_states=encoder_outputs.hidden_states,
             attentions=encoder_outputs.attentions,
             cross_attentions=encoder_outputs.cross_attentions,
@@ -437,6 +465,8 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
         encoder_hidden_states=None,
         encoder_attention_mask=None,
         labels=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
         return_dict=None,
@@ -455,6 +485,15 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
             Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
             ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are
             ignored (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``
+        past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`
+            (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`
+            instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.
+        use_cache (:obj:`bool`, `optional`):
+            If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up
+            decoding (see :obj:`past_key_values`).
 
         Returns:
 
@@ -474,6 +513,8 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
             >>> prediction_logits = outputs.logits
         """
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
 
         outputs = self.bert(
             input_ids,
@@ -483,6 +524,8 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
             inputs_embeds=inputs_embeds,
             encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -506,16 +549,26 @@ class BertGenerationDecoder(BertGenerationPreTrainedModel):
         return CausalLMOutputWithCrossAttentions(
             loss=lm_loss,
             logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
             cross_attentions=outputs.cross_attentions,
         )
 
-    def prepare_inputs_for_generation(self, input_ids, attention_mask=None, **model_kwargs):
+    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, **model_kwargs):
         input_shape = input_ids.shape
-
         # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
         if attention_mask is None:
             attention_mask = input_ids.new_ones(input_shape)
 
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
         return {"input_ids": input_ids, "attention_mask": attention_mask}
+
+    def _reorder_cache(self, past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past

src/transformers/models/electra/modeling_electra.py

@@ -33,6 +33,7 @@ from ...file_utils import (
 )
 from ...modeling_outputs import (
     BaseModelOutputWithCrossAttentions,
+    BaseModelOutputWithPastAndCrossAttentions,
     MaskedLMOutput,
     MultipleChoiceModelOutput,
     QuestionAnsweringModelOutput,
@@ -168,7 +169,9 @@ class ElectraEmbeddings(nn.Module):
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
 
     # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.forward
-    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None):
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
         if input_ids is not None:
             input_shape = input_ids.size()
         else:
@@ -177,7 +180,7 @@ class ElectraEmbeddings(nn.Module):
         seq_length = input_shape[1]
 
         if position_ids is None:
-            position_ids = self.position_ids[:, :seq_length]
+            position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length]
 
         if token_type_ids is None:
             token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
@@ -219,6 +222,8 @@ class ElectraSelfAttention(nn.Module):
             self.max_position_embeddings = config.max_position_embeddings
             self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
 
+        self.is_decoder = config.is_decoder
+
     def transpose_for_scores(self, x):
         new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
         x = x.view(*new_x_shape)
@@ -231,6 +236,7 @@ class ElectraSelfAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         mixed_query_layer = self.query(hidden_states)
@@ -238,17 +244,37 @@ class ElectraSelfAttention(nn.Module):
         # If this is instantiated as a cross-attention module, the keys
         # and values come from an encoder; the attention mask needs to be
         # such that the encoder's padding tokens are not attended to.
-        if encoder_hidden_states is not None:
-            mixed_key_layer = self.key(encoder_hidden_states)
-            mixed_value_layer = self.value(encoder_hidden_states)
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
             attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
         else:
-            mixed_key_layer = self.key(hidden_states)
-            mixed_value_layer = self.value(hidden_states)
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
-        key_layer = self.transpose_for_scores(mixed_key_layer)
-        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
@@ -292,6 +318,9 @@ class ElectraSelfAttention(nn.Module):
         context_layer = context_layer.view(*new_context_layer_shape)
 
         outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
         return outputs
 
 
@@ -343,6 +372,7 @@ class ElectraAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         self_outputs = self.self(
@@ -351,6 +381,7 @@ class ElectraAttention(nn.Module):
             head_mask,
             encoder_hidden_states,
             encoder_attention_mask,
+            past_key_value,
             output_attentions,
         )
         attention_output = self.output(self_outputs[0], hidden_states)
@@ -411,36 +442,60 @@ class ElectraLayer(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
         self_attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             head_mask,
             output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
         )
         attention_output = self_attention_outputs[0]
-        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
 
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
         if self.is_decoder and encoder_hidden_states is not None:
             assert hasattr(
                 self, "crossattention"
             ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
             cross_attention_outputs = self.crossattention(
                 attention_output,
                 attention_mask,
                 head_mask,
                 encoder_hidden_states,
                 encoder_attention_mask,
+                cross_attn_past_key_value,
                 output_attentions,
             )
             attention_output = cross_attention_outputs[0]
-            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
 
         layer_output = apply_chunking_to_forward(
             self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
         )
         outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
         return outputs
 
     def feed_forward_chunk(self, attention_output):
@@ -463,6 +518,8 @@ class ElectraEncoder(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=False,
         output_hidden_states=False,
         return_dict=True,
@@ -470,17 +527,19 @@ class ElectraEncoder(nn.Module):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
-
+            past_key_value = past_key_values[i] if past_key_values is not None else None
             if getattr(self.config, "gradient_checkpointing", False):
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
+                        return module(*inputs, past_key_value, output_attentions)
 
                     return custom_forward
 
@@ -499,9 +558,13 @@ class ElectraEncoder(nn.Module):
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
                     output_attentions,
                 )
+
             hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
             if output_attentions:
                 all_self_attentions = all_self_attentions + (layer_outputs[1],)
                 if self.config.add_cross_attention:
@@ -513,11 +576,18 @@ class ElectraEncoder(nn.Module):
         if not return_dict:
             return tuple(
                 v
-                for v in [hidden_states, all_hidden_states, all_self_attentions, all_cross_attentions]
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
                 if v is not None
             )
-        return BaseModelOutputWithCrossAttentions(
+        return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
             hidden_states=all_hidden_states,
             attentions=all_self_attentions,
             cross_attentions=all_cross_attentions,

src/transformers/models/encoder_decoder/modeling_encoder_decoder.py

@@ -345,11 +345,11 @@ class EncoderDecoderModel(PreTrainedModel):
         decoder_input_ids=None,
         decoder_attention_mask=None,
         encoder_outputs=None,
-        past_key_values=None,  # TODO: (PVP) implement :obj:`use_cache`
+        past_key_values=None,
         inputs_embeds=None,
         decoder_inputs_embeds=None,
         labels=None,
-        use_cache=None,  # TODO: (PVP) implement :obj:`use_cache`
+        use_cache=None,
         output_attentions=None,
         output_hidden_states=None,
         return_dict=None,
@@ -413,18 +413,19 @@ class EncoderDecoderModel(PreTrainedModel):
             labels=labels,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
+            use_cache=use_cache,
+            past_key_values=past_key_values,
             return_dict=return_dict,
             **kwargs_decoder,
         )
 
-        # TODO(PVP): currently it is not possible to use `past`
         if not return_dict:
             return decoder_outputs + encoder_outputs
 
         return Seq2SeqLMOutput(
             loss=decoder_outputs.loss,
             logits=decoder_outputs.logits,
-            past_key_values=None,  # TODO(PVP) - need to implement cache for BERT, etc... before this works
+            past_key_values=decoder_outputs.past_key_values,
             decoder_hidden_states=decoder_outputs.hidden_states,
             decoder_attentions=decoder_outputs.attentions,
             cross_attentions=decoder_outputs.cross_attentions,
@@ -433,24 +434,19 @@ class EncoderDecoderModel(PreTrainedModel):
             encoder_attentions=encoder_outputs.attentions,
         )
 
-    def prepare_inputs_for_generation(self, input_ids, past=None, attention_mask=None, encoder_outputs=None, **kwargs):
-        decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids)
+    def prepare_inputs_for_generation(
+        self, input_ids, past=None, attention_mask=None, use_cache=None, encoder_outputs=None, **kwargs
+    ):
+        decoder_inputs = self.decoder.prepare_inputs_for_generation(input_ids, past=past)
         decoder_attention_mask = decoder_inputs["attention_mask"] if "attention_mask" in decoder_inputs else None
         input_dict = {
             "attention_mask": attention_mask,
             "decoder_attention_mask": decoder_attention_mask,
             "decoder_input_ids": decoder_inputs["input_ids"],
             "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "use_cache": use_cache,
         }
-
-        # Ideally all models should have a :obj:`use_cache`
-        # leave following to ifs until all have it implemented
-        if "use_cache" in decoder_inputs:
-            input_dict["decoder_use_cache"] = decoder_inputs["use_cache"]
-
-        if "past_key_values" in decoder_inputs:
-            input_dict["past_key_values"] = decoder_inputs["past_key_values"]
-
         return input_dict
 
     def _reorder_cache(self, past, beam_idx):

src/transformers/models/gpt2/modeling_gpt2.py

@@ -33,7 +33,7 @@ from ...file_utils import (
 )
 from ...modeling_outputs import (
     BaseModelOutputWithPastAndCrossAttentions,
-    CausalLMOutputWithPastAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
     SequenceClassifierOutputWithPast,
 )
 from ...modeling_utils import (
@@ -851,7 +851,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
         checkpoint="gpt2",
-        output_type=CausalLMOutputWithPastAndCrossAttentions,
+        output_type=CausalLMOutputWithCrossAttentions,
         config_class=_CONFIG_FOR_DOC,
     )
     def forward(
@@ -916,7 +916,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
             output = (lm_logits,) + transformer_outputs[1:]
             return ((loss,) + output) if loss is not None else output
 
-        return CausalLMOutputWithPastAndCrossAttentions(
+        return CausalLMOutputWithCrossAttentions(
             loss=loss,
             logits=lm_logits,
             past_key_values=transformer_outputs.past_key_values,

src/transformers/models/layoutlm/modeling_layoutlm.py

@@ -24,7 +24,7 @@ from torch.nn import CrossEntropyLoss
 from ...activations import ACT2FN
 from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
 from ...modeling_outputs import (
-    BaseModelOutputWithCrossAttentions,
+    BaseModelOutputWithPastAndCrossAttentions,
     BaseModelOutputWithPoolingAndCrossAttentions,
     MaskedLMOutput,
     TokenClassifierOutput,
@@ -151,6 +151,8 @@ class LayoutLMSelfAttention(nn.Module):
             self.max_position_embeddings = config.max_position_embeddings
             self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
 
+        self.is_decoder = config.is_decoder
+
     def transpose_for_scores(self, x):
         new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
         x = x.view(*new_x_shape)
@@ -163,6 +165,7 @@ class LayoutLMSelfAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         mixed_query_layer = self.query(hidden_states)
@@ -170,17 +173,37 @@ class LayoutLMSelfAttention(nn.Module):
         # If this is instantiated as a cross-attention module, the keys
         # and values come from an encoder; the attention mask needs to be
         # such that the encoder's padding tokens are not attended to.
-        if encoder_hidden_states is not None:
-            mixed_key_layer = self.key(encoder_hidden_states)
-            mixed_value_layer = self.value(encoder_hidden_states)
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
             attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
         else:
-            mixed_key_layer = self.key(hidden_states)
-            mixed_value_layer = self.value(hidden_states)
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
-        key_layer = self.transpose_for_scores(mixed_key_layer)
-        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
@@ -224,6 +247,9 @@ class LayoutLMSelfAttention(nn.Module):
         context_layer = context_layer.view(*new_context_layer_shape)
 
         outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
         return outputs
 
 
@@ -275,6 +301,7 @@ class LayoutLMAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         self_outputs = self.self(
@@ -283,6 +310,7 @@ class LayoutLMAttention(nn.Module):
             head_mask,
             encoder_hidden_states,
             encoder_attention_mask,
+            past_key_value,
             output_attentions,
         )
         attention_output = self.output(self_outputs[0], hidden_states)
@@ -343,36 +371,60 @@ class LayoutLMLayer(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
         self_attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             head_mask,
             output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
         )
         attention_output = self_attention_outputs[0]
-        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
 
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
         if self.is_decoder and encoder_hidden_states is not None:
             assert hasattr(
                 self, "crossattention"
             ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
             cross_attention_outputs = self.crossattention(
                 attention_output,
                 attention_mask,
                 head_mask,
                 encoder_hidden_states,
                 encoder_attention_mask,
+                cross_attn_past_key_value,
                 output_attentions,
             )
             attention_output = cross_attention_outputs[0]
-            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
 
         layer_output = apply_chunking_to_forward(
             self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
         )
         outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
         return outputs
 
     def feed_forward_chunk(self, attention_output):
@@ -395,6 +447,8 @@ class LayoutLMEncoder(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=False,
         output_hidden_states=False,
         return_dict=True,
@@ -402,17 +456,19 @@ class LayoutLMEncoder(nn.Module):
         all_hidden_states = () if output_hidden_states else None
         all_self_attentions = () if output_attentions else None
         all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
         for i, layer_module in enumerate(self.layer):
             if output_hidden_states:
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer_head_mask = head_mask[i] if head_mask is not None else None
-
+            past_key_value = past_key_values[i] if past_key_values is not None else None
             if getattr(self.config, "gradient_checkpointing", False):
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
+                        return module(*inputs, past_key_value, output_attentions)
 
                     return custom_forward
 
@@ -431,9 +487,13 @@ class LayoutLMEncoder(nn.Module):
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_value,
                     output_attentions,
                 )
+
             hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
             if output_attentions:
                 all_self_attentions = all_self_attentions + (layer_outputs[1],)
                 if self.config.add_cross_attention:
@@ -445,11 +505,18 @@ class LayoutLMEncoder(nn.Module):
         if not return_dict:
             return tuple(
                 v
-                for v in [hidden_states, all_hidden_states, all_self_attentions, all_cross_attentions]
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
                 if v is not None
             )
-        return BaseModelOutputWithCrossAttentions(
+        return BaseModelOutputWithPastAndCrossAttentions(
             last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
             hidden_states=all_hidden_states,
             attentions=all_self_attentions,
             cross_attentions=all_cross_attentions,

src/transformers/models/longformer/modeling_longformer.py

@@ -424,7 +424,6 @@ def _compute_global_attention_mask(input_ids, sep_token_id, before_sep_token=Tru
     return attention_mask
 
 
-# Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids
 def create_position_ids_from_input_ids(input_ids, padding_idx):
     """
     Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols

src/transformers/models/tapas/modeling_tapas.py

@@ -347,6 +347,7 @@ class TapasSelfAttention(nn.Module):
         self.value = nn.Linear(config.hidden_size, self.all_head_size)
 
         self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.is_decoder = config.is_decoder
 
     def transpose_for_scores(self, x):
         new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
@@ -360,6 +361,7 @@ class TapasSelfAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         mixed_query_layer = self.query(hidden_states)
@@ -367,17 +369,30 @@ class TapasSelfAttention(nn.Module):
         # If this is instantiated as a cross-attention module, the keys
         # and values come from an encoder; the attention mask needs to be
         # such that the encoder's padding tokens are not attended to.
-        if encoder_hidden_states is not None:
-            mixed_key_layer = self.key(encoder_hidden_states)
-            mixed_value_layer = self.value(encoder_hidden_states)
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
             attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
         else:
-            mixed_key_layer = self.key(hidden_states)
-            mixed_value_layer = self.value(hidden_states)
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
 
         query_layer = self.transpose_for_scores(mixed_query_layer)
-        key_layer = self.transpose_for_scores(mixed_key_layer)
-        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        if self.is_decoder:
+            past_key_value = (key_layer, value_layer)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
         attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
@@ -404,6 +419,8 @@ class TapasSelfAttention(nn.Module):
         context_layer = context_layer.view(*new_context_layer_shape)
 
         outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
         return outputs
 
 
@@ -455,6 +472,7 @@ class TapasAttention(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
         self_outputs = self.self(
@@ -463,6 +481,7 @@ class TapasAttention(nn.Module):
             head_mask,
             encoder_hidden_states,
             encoder_attention_mask,
+            past_key_value,
             output_attentions,
         )
         attention_output = self.output(self_outputs[0], hidden_states)
@@ -523,36 +542,60 @@ class TapasLayer(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_value=None,
         output_attentions=False,
     ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
         self_attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             head_mask,
             output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
         )
         attention_output = self_attention_outputs[0]
-        outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
 
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
         if self.is_decoder and encoder_hidden_states is not None:
             assert hasattr(
                 self, "crossattention"
             ), f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`"
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
             cross_attention_outputs = self.crossattention(
                 attention_output,
                 attention_mask,
                 head_mask,
                 encoder_hidden_states,
                 encoder_attention_mask,
+                cross_attn_past_key_value,
                 output_attentions,
             )
             attention_output = cross_attention_outputs[0]
-            outputs = outputs + cross_attention_outputs[1:]  # add cross attentions if we output attention weights
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
 
         layer_output = apply_chunking_to_forward(
             self.feed_forward_chunk, self.chunk_size_feed_forward, self.seq_len_dim, attention_output
         )
         outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
         return outputs
 
     def feed_forward_chunk(self, attention_output):
@@ -574,6 +617,8 @@ class TapasEncoder(nn.Module):
         head_mask=None,
         encoder_hidden_states=None,
         encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
         output_attentions=False,
         output_hidden_states=False,
         return_dict=True,
@@ -590,7 +635,7 @@ class TapasEncoder(nn.Module):
 
                 def create_custom_forward(module):
                     def custom_forward(*inputs):
-                        return module(*inputs, output_attentions)
+                        return module(*inputs, past_key_values, output_attentions)
 
                     return custom_forward
 
@@ -609,6 +654,7 @@ class TapasEncoder(nn.Module):
                     layer_head_mask,
                     encoder_hidden_states,
                     encoder_attention_mask,
+                    past_key_values,
                     output_attentions,
                 )
             hidden_states = layer_outputs[0]

tests/test_modeling_bart.py

@@ -150,7 +150,7 @@ class BartModelTester:
         input_ids = inputs_dict["input_ids"]
 
         # first forward pass
-        outputs = model(input_ids, use_cache=True)
+        outputs = model(input_ids, attention_mask=inputs_dict["attention_mask"], use_cache=True)
 
         output, past_key_values = outputs.to_tuple()
 

tests/test_modeling_bert.py

@@ -260,6 +260,66 @@ class BertModelTester:
         )
         self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
 
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = BertLMHeadModel(config=config).to(torch_device).eval()
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
     def create_and_check_for_next_sequence_prediction(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
@@ -454,6 +514,10 @@ class BertModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
         self.model_tester.create_and_check_model_for_causal_lm_as_decoder(*config_and_inputs)
 
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
     def test_for_multiple_choice(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)

tests/test_modeling_bert_generation.py

@@ -25,6 +25,8 @@ from .test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, r
 
 
 if is_torch_available():
+    import torch
+
     from transformers import BertGenerationConfig, BertGenerationDecoder, BertGenerationEncoder
 
 
@@ -156,6 +158,64 @@ class BertGenerationEncoderTester:
         )
         self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
 
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        input_mask,
+        token_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+        **kwargs,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = BertGenerationDecoder(config=config).to(torch_device).eval()
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
     def create_and_check_for_causal_lm(
         self,
         config,
@@ -203,6 +263,10 @@ class BertGenerationEncoderTest(ModelTesterMixin, GenerationTesterMixin, unittes
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
         self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
 
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
     def test_model_as_decoder_with_default_input_mask(self):
         # This regression test was failing with PyTorch < 1.3
         (

tests/test_modeling_roberta.py

@@ -198,6 +198,74 @@ class RobertaModelTester:
         result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
         self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
 
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = RobertaForCausalLM(config=config).to(torch_device).eval()
+
+        # make sure that ids don't start with pad token
+        mask = input_ids.ne(config.pad_token_id).long()
+        input_ids = input_ids * mask
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+
+        # make sure that ids don't start with pad token
+        mask = next_tokens.ne(config.pad_token_id).long()
+        next_tokens = next_tokens * mask
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
     def create_and_check_for_masked_lm(
         self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
     ):
@@ -337,6 +405,10 @@ class RobertaModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCas
         config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
         self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
 
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
     def test_for_masked_lm(self):
         config_and_inputs = self.model_tester.prepare_config_and_inputs()
         self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)