Fix `input_embeds` docstring in encoder-decoder architectures (#28168)

src/transformers/models/bart/modeling_bart.py

@@ -1011,10 +1011,11 @@ BART_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -1331,11 +1332,11 @@ class BartDecoder(BartPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/bart/modeling_tf_bart.py

@@ -715,6 +715,10 @@ BART_INPUTS_DOCSTRING = r"""
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
             `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         use_cache (`bool`, *optional*, defaults to `True`):
             If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
             `past_key_values`). Set to `False` during training, `True` during generation
@@ -990,11 +994,11 @@ class TFBartDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.tTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/bigbird_pegasus/modeling_bigbird_pegasus.py

@@ -1678,10 +1678,11 @@ BIGBIRD_PEGASUS_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -2136,11 +2137,11 @@ class BigBirdPegasusDecoder(BigBirdPegasusPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/biogpt/modeling_biogpt.py

@@ -396,10 +396,11 @@ BIOGPT_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         use_cache (`bool`, *optional*):
             If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
             `past_key_values`).

src/transformers/models/blenderbot/modeling_blenderbot.py

@@ -589,10 +589,11 @@ BLENDERBOT_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -892,11 +893,11 @@ class BlenderbotDecoder(BlenderbotPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/blenderbot/modeling_tf_blenderbot.py

@@ -949,11 +949,11 @@ class TFBlenderbotDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value

src/transformers/models/blenderbot_small/modeling_blenderbot_small.py

@@ -589,10 +589,11 @@ BLENDERBOT_SMALL_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -889,11 +890,11 @@ class BlenderbotSmallDecoder(BlenderbotSmallPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/blenderbot_small/modeling_tf_blenderbot_small.py

@@ -957,11 +957,11 @@ class TFBlenderbotSmallDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value

src/transformers/models/led/modeling_led.py

@@ -2021,11 +2021,11 @@ class LEDDecoder(LEDPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/m2m_100/modeling_m2m_100.py

@@ -630,10 +630,11 @@ M2M_100_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -931,11 +932,11 @@ class M2M100Decoder(M2M100PreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/marian/modeling_tf_marian.py

@@ -688,6 +688,10 @@ MARIAN_INPUTS_DOCSTRING = r"""
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
             `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         use_cache (`bool`, *optional*, defaults to `True`):
             If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
             `past_key_values`). Set to `False` during training, `True` during generation
@@ -975,11 +979,11 @@ class TFMarianDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value

src/transformers/models/mbart/modeling_mbart.py

@@ -876,10 +876,11 @@ MBART_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -1191,11 +1192,11 @@ class MBartDecoder(MBartPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/mbart/modeling_tf_mbart.py

@@ -636,6 +636,10 @@ MBART_INPUTS_DOCSTRING = r"""
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
             `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         use_cache (`bool`, *optional*, defaults to `True`):
             If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
             `past_key_values`). Set to `False` during training, `True` during generation
@@ -981,11 +985,11 @@ class TFMBartDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value

src/transformers/models/musicgen/modeling_musicgen.py

@@ -539,10 +539,11 @@ MUSICGEN_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -626,10 +627,11 @@ MUSICGEN_DECODER_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+            This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
         output_attentions (`bool`, *optional*):
             Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
             tensors for more detail.

src/transformers/models/mvp/modeling_mvp.py

@@ -629,10 +629,11 @@ MVP_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -1067,11 +1068,11 @@ class MvpDecoder(MvpPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/nllb_moe/modeling_nllb_moe.py

@@ -943,10 +943,11 @@ NLLB_MOE_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -1271,11 +1272,11 @@ class NllbMoeDecoder(NllbMoePreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/pegasus/modeling_pegasus.py

@@ -584,10 +584,11 @@ PEGASUS_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
@@ -946,11 +947,11 @@ class PegasusDecoder(PegasusPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.

src/transformers/models/pegasus/modeling_tf_pegasus.py

@@ -688,6 +688,10 @@ PEGASUS_INPUTS_DOCSTRING = r"""
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
             `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         use_cache (`bool`, *optional*, defaults to `True`):
             If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
             `past_key_values`). Set to `False` during training, `True` during generation output_attentions (`bool`,
@@ -985,11 +989,11 @@ class TFPegasusDecoder(tf.keras.layers.Layer):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
-                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value

src/transformers/models/pegasus_x/modeling_pegasus_x.py

@@ -840,10 +840,11 @@ PEGASUS_X_INPUTS_DOCSTRING = r"""
 
             If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
             don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
-            `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape
-            `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `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.
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `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.
         decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
             Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
             representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be

src/transformers/models/plbart/modeling_plbart.py

@@ -938,11 +938,11 @@ class PLBartDecoder(PLBartPreTrainedModel):
 
                 If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
                 that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
-                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of
-                shape `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
-                `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.
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
             output_attentions (`bool`, *optional*):
                 Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                 returned tensors for more detail.