[GPT2, CTRL] Allow input of input_ids and past of variable length (#4581)

* revert convenience  method

* clean docs a bit

src/transformers/modeling_ctrl.py

@@ -208,9 +208,11 @@ CTRL_START_DOCSTRING = r"""
 
 CTRL_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only input_ids that do not have their past calculated should be passed as input_ids.
 
             Indices can be obtained using :class:`transformers.CTRLTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -220,9 +222,7 @@ CTRL_INPUTS_DOCSTRING = r"""
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids`
+            The input_ids which have their past given to this model should not be passed as input ids as they have already been computed.
-            (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)`
-            instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -233,7 +233,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -246,7 +245,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
         input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
-            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.
             If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
@@ -344,16 +342,6 @@ class CTRLModel(CTRLPreTrainedModel):
 
         """
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         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:

src/transformers/modeling_gpt2.py

@@ -286,9 +286,11 @@ GPT2_START_DOCSTRING = r"""
 
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`.
 
             Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -299,7 +301,7 @@ GPT2_INPUTS_DOCSTRING = r"""
         past (:obj:`List[torch.FloatTensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids` (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)` instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
+            The `input_ids` which have their past given to this model should not be passed as `input_ids` as they have already been computed.
         attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -311,8 +313,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
-
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Indices of positions of each input sequence tokens in the position embeddings.
@@ -324,7 +324,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Mask values selected in ``[0, 1]``:
             :obj:`1` indicates the head is **not masked**, :obj:`0` indicates the head is **masked**.
         input_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`, defaults to :obj:`None`):
-            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.
             If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
@@ -410,16 +409,6 @@ class GPT2Model(GPT2PreTrainedModel):
 
         """
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         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:

src/transformers/modeling_tf_ctrl.py

@@ -411,9 +411,12 @@ CTRL_START_DOCSTRING = r"""
 
 CTRL_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
+
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only input_ids that do not have their past calculated should be passed as input_ids (see `past`).
 
             Indices can be obtained using :class:`transformers.CTRLTokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -423,9 +426,8 @@ CTRL_INPUTS_DOCSTRING = r"""
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
             (see `past` output below). Can be used to speed up sequential decoding.
-            If `past` is used, the user can optionally input only the last `input_ids`
+            The token ids which have their past given to this model
-            (those that don't have their past given to this model) of shape :obj:`(batch_size, 1)`
+            should not be passed as input ids as they have already been computed.
-            instead of all `input_ids` of shape :obj:`(batch_size, sequence_length)`.
         attention_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -436,7 +438,6 @@ CTRL_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -452,7 +453,6 @@ CTRL_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.
-            If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
         use_cache (:obj:`bool`):
             If `use_cache` is True, `past` key value states are returned and
             can be used to speed up decoding (see `past`). Defaults to `True`.

src/transformers/modeling_tf_gpt2.py

@@ -284,16 +284,6 @@ class TFGPT2MainLayer(tf.keras.layers.Layer):
         else:
             input_ids = inputs
 
-        # If using past key value states, only the last tokens
-        # should be given as an input
-        if past is not None:
-            if input_ids is not None:
-                input_ids = input_ids[:, -1:]
-            if inputs_embeds is not None:
-                inputs_embeds = inputs_embeds[:, -1:]
-            if token_type_ids is not None:
-                token_type_ids = token_type_ids[:, -1:]
-
         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:
@@ -431,9 +421,11 @@ GPT2_START_DOCSTRING = r"""
 
 GPT2_INPUTS_DOCSTRING = r"""
     Args:
-        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, sequence_length)`):
+        input_ids (:obj:`Numpy array` or :obj:`tf.Tensor` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length`` if ``past`` is ``None`` else ``past[0].shape[-2]`` (``sequence_length`` of input past key value states).
             Indices of input sequence tokens in the vocabulary.
-            If `past` is used, optionally only the last `input_ids` have to be input (see `past`).
+
+            If `past` is used, only `input_ids` that do not have their past calculated should be passed as `input_ids`.
 
             Indices can be obtained using :class:`transformers.GPT2Tokenizer`.
             See :func:`transformers.PreTrainedTokenizer.encode` and
@@ -442,8 +434,9 @@ GPT2_INPUTS_DOCSTRING = r"""
             `What are input IDs? <../glossary.html#input-ids>`__
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers`):
             Contains pre-computed hidden-states (key and values in the attention blocks) as computed by the model
-            (see `past` output below). Can be used to speed up sequential decoding. The token ids which have their past given to this model
+            (see `past` output below). Can be used to speed up sequential decoding.
-            should not be passed as input ids as they have already been computed.
+            The token ids which have their past given to this model
+            should not be passed as `input_ids` as they have already been computed.
         attention_mask (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
             Mask to avoid performing attention on padding token indices.
             Mask values selected in ``[0, 1]``:
@@ -454,7 +447,6 @@ GPT2_INPUTS_DOCSTRING = r"""
             Segment token indices to indicate first and second portions of the inputs.
             Indices are selected in ``[0, 1]``: ``0`` corresponds to a `sentence A` token, ``1``
             corresponds to a `sentence B` token
-            If `past` is used, optionally only the last `token_type_ids` have to be input (see `past`).
 
             `What are token type IDs? <../glossary.html#token-type-ids>`_
         position_ids (:obj:`tf.Tensor` or :obj:`Numpy array` of shape :obj:`(batch_size, sequence_length)`, `optional`, defaults to :obj:`None`):
@@ -470,7 +462,6 @@ GPT2_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.
-            If `past` is used, optionally only the last `input_embeds` have to be input (see `past`).
         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.
@@ -639,7 +630,7 @@ class TFGPT2DoubleHeadsModel(TFGPT2PreTrainedModel):
         past (:obj:`List[tf.Tensor]` of length :obj:`config.n_layers` with each tensor of shape :obj:`(2, batch_size, num_heads, sequence_length, embed_size_per_head)`):
             Contains pre-computed hidden-states (key and values in the attention blocks).
             Can be used (see `past` input) to speed up sequential decoding. The token ids which have their past given to this model
-            should not be passed as input ids as they have already been computed.
+            should not be passed as `input_ids` as they have already been computed.
         hidden_states (:obj:`tuple(tf.Tensor)`, `optional`, returned when ``config.output_hidden_states=True``):
             Tuple of :obj:`tf.Tensor` (one for the output of the embeddings + one for the output of each layer)
             of shape :obj:`(batch_size, sequence_length, hidden_size)`.