modeling_blenderbot.py 74.6 KB
Newer Older
Sam Shleifer's avatar
Sam Shleifer committed
1
# coding=utf-8
2
# Copyright 2021 The Facebook, Inc. and The HuggingFace Inc. team. All rights reserved.
Sam Shleifer's avatar
Sam Shleifer committed
3
#
4
# Licensed under the Apache License, Version 2.0 (the "License");
Sam Shleifer's avatar
Sam Shleifer committed
5
6
7
8
9
10
11
12
13
14
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
Sylvain Gugger's avatar
Sylvain Gugger committed
15
""" PyTorch Blenderbot model."""
16
17


18
import copy
19
20
21
22
import math
import os
import random
import warnings
23
from typing import List, Optional, Tuple, Union
Sam Shleifer's avatar
Sam Shleifer committed
24
25

import torch
26
import torch.utils.checkpoint
27
28
from torch import nn
from torch.nn import CrossEntropyLoss
Sam Shleifer's avatar
Sam Shleifer committed
29

30
31
32
33
from ...activations import ACT2FN
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
34
    CausalLMOutputWithCrossAttentions,
35
36
37
38
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
)
from ...modeling_utils import PreTrainedModel
39
40
41
42
43
44
45
from ...utils import (
    add_end_docstrings,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    logging,
    replace_return_docstrings,
)
46
from ..blenderbot_small import BlenderbotSmallForConditionalGeneration, BlenderbotSmallModel
Sam Shleifer's avatar
Sam Shleifer committed
47
48
49
from .configuration_blenderbot import BlenderbotConfig


50
51
52
53
logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "BlenderbotConfig"
_TOKENIZER_FOR_DOC = "BlenderbotTokenizer"
54
_CHECKPOINT_FOR_DOC = "facebook/blenderbot-400M-distill"
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71


BLENDERBOT_PRETRAINED_MODEL_ARCHIVE_LIST = [
    "facebook/blenderbot-3B",
    # See all Blenderbot models at https://huggingface.co/models?filter=blenderbot
]


# Copied from transformers.models.bart.modeling_bart.shift_tokens_right
def shift_tokens_right(input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

72
73
    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
74
75
76
77
78
79
80
81
82
83
84
85
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids


# Copied from transformers.models.bart.modeling_bart._make_causal_mask
def _make_causal_mask(input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0):
    """
    Make causal mask used for bi-directional self-attention.
    """
    bsz, tgt_len = input_ids_shape
Yih-Dar's avatar
Yih-Dar committed
86
    mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min))
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
    mask_cond = torch.arange(mask.size(-1))
    mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)
    mask = mask.to(dtype)

    if past_key_values_length > 0:
        mask = torch.cat([torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1)
    return mask[None, None, :, :].expand(bsz, 1, tgt_len, tgt_len + past_key_values_length)


# Copied from transformers.models.bart.modeling_bart._expand_mask
def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
    """
    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
    """
    bsz, src_len = mask.size()
    tgt_len = tgt_len if tgt_len is not None else src_len

    expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)

    inverted_mask = 1.0 - expanded_mask

108
    return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
109
110
111
112
113
114
115


class BlenderbotLearnedPositionalEmbedding(nn.Embedding):
    """
    This module learns positional embeddings up to a fixed maximum size.
    """

116
117
    def __init__(self, num_embeddings: int, embedding_dim: int):
        super().__init__(num_embeddings, embedding_dim)
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144

    def forward(self, input_ids_shape: torch.Size, past_key_values_length: int = 0):
        """`input_ids_shape` is expected to be [bsz x seqlen]."""
        bsz, seq_len = input_ids_shape[:2]
        positions = torch.arange(
            past_key_values_length, past_key_values_length + seq_len, dtype=torch.long, device=self.weight.device
        )
        return super().forward(positions)


# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->Blenderbot
class BlenderbotAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        embed_dim: int,
        num_heads: int,
        dropout: float = 0.0,
        is_decoder: bool = False,
        bias: bool = True,
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.dropout = dropout
        self.head_dim = embed_dim // num_heads
145
146
147
148
149
150

        if (self.head_dim * num_heads) != self.embed_dim:
            raise ValueError(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
                f" and `num_heads`: {num_heads})."
            )
151
        self.scaling = self.head_dim**-0.5
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
        self.is_decoder = is_decoder

        self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)
        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)

    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()

    def forward(
        self,
        hidden_states: torch.Tensor,
        key_value_states: Optional[torch.Tensor] = None,
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        attention_mask: Optional[torch.Tensor] = None,
168
        layer_head_mask: Optional[torch.Tensor] = None,
169
170
171
172
173
174
175
        output_attentions: bool = False,
    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
        """Input shape: Batch x Time x Channel"""

        # if key_value_states are provided this layer is used as a cross-attention layer
        # for the decoder
        is_cross_attention = key_value_states is not None
176
177

        bsz, tgt_len, _ = hidden_states.size()
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218

        # get query proj
        query_states = self.q_proj(hidden_states) * self.scaling
        # get key, value proj
        if is_cross_attention and past_key_value is not None:
            # reuse k,v, cross_attentions
            key_states = past_key_value[0]
            value_states = past_key_value[1]
        elif is_cross_attention:
            # cross_attentions
            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
        elif past_key_value is not None:
            # reuse k, v, self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
            key_states = torch.cat([past_key_value[0], key_states], dim=2)
            value_states = torch.cat([past_key_value[1], value_states], dim=2)
        else:
            # self_attention
            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)

        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_states, value_states)

        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
        query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
        key_states = key_states.view(*proj_shape)
        value_states = value_states.view(*proj_shape)

        src_len = key_states.size(1)
        attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))

219
220
        if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
            raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
221
222
                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
                f" {attn_weights.size()}"
223
            )
224
225

        if attention_mask is not None:
226
227
228
229
            if attention_mask.size() != (bsz, 1, tgt_len, src_len):
                raise ValueError(
                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
                )
230
231
232
            attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

233
        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
234

235
        if layer_head_mask is not None:
236
237
            if layer_head_mask.size() != (self.num_heads,):
                raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
238
239
                    f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
                    f" {layer_head_mask.size()}"
240
                )
241
242
243
            attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)

244
        if output_attentions:
245
            # this operation is a bit awkward, but it's required to
246
            # make sure that attn_weights keeps its gradient.
247
            # In order to do so, attn_weights have to be reshaped
248
249
250
251
252
253
            # twice and have to be reused in the following
            attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
            attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
        else:
            attn_weights_reshaped = None

254
        attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
255
256
257

        attn_output = torch.bmm(attn_probs, value_states)

258
259
        if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
            raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
260
261
                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
                f" {attn_output.size()}"
262
            )
263

264
265
        attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
        attn_output = attn_output.transpose(1, 2)
266
267
268
269

        # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
        # partitioned aross GPUs when using tensor-parallelism.
        attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)
270
271
272
273

        attn_output = self.out_proj(attn_output)

        return attn_output, attn_weights_reshaped, past_key_value
Sam Shleifer's avatar
Sam Shleifer committed
274

275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293

# Copied from transformers.models.mbart.modeling_mbart.MBartEncoderLayer with MBart->Blenderbot
class BlenderbotEncoderLayer(nn.Module):
    def __init__(self, config: BlenderbotConfig):
        super().__init__()
        self.embed_dim = config.d_model
        self.self_attn = BlenderbotAttention(
            embed_dim=self.embed_dim,
            num_heads=config.encoder_attention_heads,
            dropout=config.attention_dropout,
        )
        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout
        self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
        self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

294
295
296
297
298
299
    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: torch.Tensor,
        layer_head_mask: torch.Tensor,
        output_attentions: bool = False,
300
    ) -> torch.Tensor:
301
302
        """
        Args:
303
304
305
306
307
308
309
            hidden_states (`torch.FloatTensor`): input to the layer of shape *(seq_len, batch, embed_dim)*
            attention_mask (`torch.FloatTensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                *(encoder_attention_heads,)*.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
310
311
312
313
314
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)
        hidden_states, attn_weights, _ = self.self_attn(
315
316
317
318
            hidden_states=hidden_states,
            attention_mask=attention_mask,
            layer_head_mask=layer_head_mask,
            output_attentions=output_attentions,
319
        )
320
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
321
322
323
324
325
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
326
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
327
        hidden_states = self.fc2(hidden_states)
328
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
329
330
        hidden_states = residual + hidden_states

331
332
333
        if hidden_states.dtype == torch.float16 and (
            torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
        ):
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (attn_weights,)

        return outputs


# Copied from transformers.models.mbart.modeling_mbart.MBartDecoderLayer with MBart->Blenderbot
class BlenderbotDecoderLayer(nn.Module):
    def __init__(self, config: BlenderbotConfig):
        super().__init__()
        self.embed_dim = config.d_model

        self.self_attn = BlenderbotAttention(
            embed_dim=self.embed_dim,
            num_heads=config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        self.dropout = config.dropout
        self.activation_fn = ACT2FN[config.activation_function]
        self.activation_dropout = config.activation_dropout

        self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.encoder_attn = BlenderbotAttention(
            self.embed_dim,
            config.decoder_attention_heads,
            dropout=config.attention_dropout,
            is_decoder=True,
        )
        self.encoder_attn_layer_norm = nn.LayerNorm(self.embed_dim)
        self.fc1 = nn.Linear(self.embed_dim, config.decoder_ffn_dim)
        self.fc2 = nn.Linear(config.decoder_ffn_dim, self.embed_dim)
        self.final_layer_norm = nn.LayerNorm(self.embed_dim)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        encoder_attention_mask: Optional[torch.Tensor] = None,
379
        layer_head_mask: Optional[torch.Tensor] = None,
380
        cross_attn_layer_head_mask: Optional[torch.Tensor] = None,
381
382
383
        past_key_value: Optional[Tuple[torch.Tensor]] = None,
        output_attentions: Optional[bool] = False,
        use_cache: Optional[bool] = True,
384
    ) -> torch.Tensor:
385
386
        """
        Args:
387
388
389
            hidden_states (`torch.FloatTensor`): input to the layer of shape *(seq_len, batch, embed_dim)*
            attention_mask (`torch.FloatTensor`): attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
Sylvain Gugger's avatar
Sylvain Gugger committed
390
391
            encoder_hidden_states (`torch.FloatTensor`):
                cross attention input to the layer of shape *(seq_len, batch, embed_dim)*
392
393
394
395
396
397
398
399
400
            encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size
                *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values.
            layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
                *(encoder_attention_heads,)*.
            cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of
                size *(decoder_attention_heads,)*.
            past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
401
402
403
404
405
406
407
408
409
410
411
412
413
                returned tensors for more detail.
        """
        residual = hidden_states
        hidden_states = self.self_attn_layer_norm(hidden_states)

        # Self Attention
        # 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
        # add present self-attn cache to positions 1,2 of present_key_value tuple
        hidden_states, self_attn_weights, present_key_value = self.self_attn(
            hidden_states=hidden_states,
            past_key_value=self_attn_past_key_value,
            attention_mask=attention_mask,
414
            layer_head_mask=layer_head_mask,
415
416
            output_attentions=output_attentions,
        )
417
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
        hidden_states = residual + hidden_states

        # Cross-Attention Block
        cross_attn_present_key_value = None
        cross_attn_weights = None
        if encoder_hidden_states is not None:
            residual = hidden_states
            hidden_states = self.encoder_attn_layer_norm(hidden_states)

            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
                hidden_states=hidden_states,
                key_value_states=encoder_hidden_states,
                attention_mask=encoder_attention_mask,
433
                layer_head_mask=cross_attn_layer_head_mask,
434
435
436
                past_key_value=cross_attn_past_key_value,
                output_attentions=output_attentions,
            )
437
            hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
438
439
440
441
442
443
444
445
446
            hidden_states = residual + hidden_states

            # add cross-attn to positions 3,4 of present_key_value tuple
            present_key_value = present_key_value + cross_attn_present_key_value

        # Fully Connected
        residual = hidden_states
        hidden_states = self.final_layer_norm(hidden_states)
        hidden_states = self.activation_fn(self.fc1(hidden_states))
447
        hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
448
        hidden_states = self.fc2(hidden_states)
449
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
        hidden_states = residual + hidden_states

        outputs = (hidden_states,)

        if output_attentions:
            outputs += (self_attn_weights, cross_attn_weights)

        if use_cache:
            outputs += (present_key_value,)

        return outputs


class BlenderbotPreTrainedModel(PreTrainedModel):
    config_class = BlenderbotConfig
    base_model_prefix = "model"
466
    supports_gradient_checkpointing = True
467
468
469
470
471
472
473
474
475
476
477
478

    def _init_weights(self, module):
        std = self.config.init_std
        if isinstance(module, nn.Linear):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

479
480
481
482
    def _set_gradient_checkpointing(self, module, value=False):
        if isinstance(module, (BlenderbotDecoder, BlenderbotEncoder)):
            module.gradient_checkpointing = value

483
484
485
486
487
488
489
490
491
492
493
494
495
    @property
    def dummy_inputs(self):
        pad_token = self.config.pad_token_id
        input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
        dummy_inputs = {
            "attention_mask": input_ids.ne(pad_token),
            "input_ids": input_ids,
            "decoder_input_ids": input_ids,
        }
        return dummy_inputs


BLENDERBOT_START_DOCSTRING = r"""
Sylvain Gugger's avatar
Sylvain Gugger committed
496
497
498
    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
    etc.)
Sam Shleifer's avatar
Sam Shleifer committed
499

Sylvain Gugger's avatar
Sylvain Gugger committed
500
501
502
    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
    and behavior.
Sam Shleifer's avatar
Sam Shleifer committed
503

504
    Parameters:
505
        config ([`BlenderbotConfig`]):
506
507
            Model configuration class with all the parameters of the model. Initializing with a config file does not
            load the weights associated with the model, only the configuration. Check out the
508
            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
Sam Shleifer's avatar
Sam Shleifer committed
509
510
"""

511
BLENDERBOT_GENERATION_EXAMPLE = r"""
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
    Conversation example:

    ```python
    >>> from transformers import BlenderbotTokenizer, BlenderbotForConditionalGeneration

    >>> mname = "facebook/blenderbot-400M-distill"
    >>> model = BlenderbotForConditionalGeneration.from_pretrained(mname)
    >>> tokenizer = BlenderbotTokenizer.from_pretrained(mname)
    >>> UTTERANCE = "My friends are cool but they eat too many carbs."
    >>> print("Human: ", UTTERANCE)
    Human:  My friends are cool but they eat too many carbs.

    >>> inputs = tokenizer([UTTERANCE], return_tensors="pt")
    >>> reply_ids = model.generate(**inputs)
    >>> print("Bot: ", tokenizer.batch_decode(reply_ids, skip_special_tokens=True)[0])
    Bot: That's unfortunate. Are they trying to lose weight or are they just trying to be healthier?

    >>> REPLY = "I'm not sure"
    >>> print("Human: ", REPLY)
    Human: I'm not sure

    >>> NEXT_UTTERANCE = (
    ...     "My friends are cool but they eat too many carbs.</s> <s>That's unfortunate. "
    ...     "Are they trying to lose weight or are they just trying to be healthier?</s> "
    ...     "<s> I'm not sure."
    ... )
    >>> inputs = tokenizer([NEXT_UTTERANCE], return_tensors="pt")
    >>> next_reply_ids = model.generate(**inputs)
    >>> print("Bot: ", tokenizer.batch_decode(next_reply_ids, skip_special_tokens=True)[0])
    Bot:   That's too bad. Have you tried encouraging them to change their eating habits?
    ```
543
544
545
546
"""

BLENDERBOT_INPUTS_DOCSTRING = r"""
    Args:
547
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
548
549
550
            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
            it.

Sylvain Gugger's avatar
Sylvain Gugger committed
551
552
            Indices can be obtained using [`BlenderbotTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.
553

554
555
556
            [What are input IDs?](../glossary#input-ids)
        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
557
558
559
560

            - 1 for tokens that are **not masked**,
            - 0 for tokens that are **masked**.

561
562
            [What are attention masks?](../glossary#attention-mask)
        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
563
564
            Indices of decoder input sequence tokens in the vocabulary.

Sylvain Gugger's avatar
Sylvain Gugger committed
565
566
            Indices can be obtained using [`BlenderbotTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.
567

568
            [What are decoder input IDs?](../glossary#decoder-input-ids)
569

570
571
572
573
            Blenderbot uses the `bos_token_id` as the starting token for `decoder_input_ids` generation. If
            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
            `past_key_values`).
        decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
574
575
            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
            be used by default.
576
577
        head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
578
579

            - 1 indicates the head is **not masked**,
580
            - 0 indicates the head is **masked**.
581

582
583
        decoder_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
584
585
586
587

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

588
        cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
589
590
            Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0,
            1]`:
591
592
593
594

            - 1 indicates the head is **not masked**,
            - 0 indicates the head is **masked**.

595
        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
596
597
598
            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
599
        past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Sylvain Gugger's avatar
Sylvain Gugger committed
600
601
602
            Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
603
604

            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
605
606
            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

Sylvain Gugger's avatar
Sylvain Gugger committed
607
608
            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
609
610
611
612
            `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.
613
614
        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
Sylvain Gugger's avatar
Sylvain Gugger committed
615
616
            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
            input (see `past_key_values`). This is useful if you want more control over how to convert
617
618
            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.

Sylvain Gugger's avatar
Sylvain Gugger committed
619
620
            If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value
            of `inputs_embeds`.
621
        use_cache (`bool`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
622
623
            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
            `past_key_values`).
624
625
        output_attentions (`bool`, *optional*):
            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
626
            tensors for more detail.
627
628
        output_hidden_states (`bool`, *optional*):
            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
629
            more detail.
630
        return_dict (`bool`, *optional*):
631
            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
632
633
634
635
636
637
"""


class BlenderbotEncoder(BlenderbotPreTrainedModel):
    """
    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
638
    [`BlenderbotEncoderLayer`].
639
640
641

    Args:
        config: BlenderbotConfig
642
        embed_tokens (nn.Embedding): output embedding
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
    """

    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)

        self.dropout = config.dropout
        self.layerdrop = config.encoder_layerdrop

        embed_dim = config.d_model
        self.padding_idx = config.pad_token_id
        self.max_source_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(embed_dim) if config.scale_embedding else 1.0

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim, self.padding_idx)

        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
            config.max_position_embeddings,
            embed_dim,
        )
        self.layers = nn.ModuleList([BlenderbotEncoderLayer(config) for _ in range(config.encoder_layers)])
        self.layer_norm = nn.LayerNorm(config.d_model)

668
        self.gradient_checkpointing = False
669
670
        # Initialize weights and apply final processing
        self.post_init()
671
672
673
674
675

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
676
        head_mask=None,
677
678
679
680
681
682
683
        inputs_embeds=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
684
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
685
686
687
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

Sylvain Gugger's avatar
Sylvain Gugger committed
688
689
                Indices can be obtained using [`BlenderbotTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.
690

691
692
693
                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
694
695
696
697

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

698
699
700
                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
701
702

                - 1 indicates the head is **not masked**,
703
                - 0 indicates the head is **masked**.
704

705
            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
706
707
708
                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.
709
710
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
711
                returned tensors for more detail.
712
713
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
714
                for more detail.
715
            return_dict (`bool`, *optional*):
716
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        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()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        embed_pos = self.embed_positions(input_shape)

        hidden_states = inputs_embeds + embed_pos
741
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
742
743
744
745
746
747
748
749

        # expand attention_mask
        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)

        encoder_states = () if output_hidden_states else None
        all_attentions = () if output_attentions else None
750
751
752

        # check if head_mask has a correct number of layers specified if desired
        if head_mask is not None:
753
754
            if head_mask.size()[0] != len(self.layers):
                raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
755
756
                    f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
                    f" {head_mask.size()[0]}."
757
                )
758
        for idx, encoder_layer in enumerate(self.layers):
759
760
761
762
763
764
765
            if output_hidden_states:
                encoder_states = encoder_states + (hidden_states,)
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):  # skip the layer
                layer_outputs = (None, None)
            else:
766
                if self.gradient_checkpointing and self.training:
767
768
769
770
771
772
773
774
775
776
777

                    def create_custom_forward(module):
                        def custom_forward(*inputs):
                            return module(*inputs, output_attentions)

                        return custom_forward

                    layer_outputs = torch.utils.checkpoint.checkpoint(
                        create_custom_forward(encoder_layer),
                        hidden_states,
                        attention_mask,
778
                        (head_mask[idx] if head_mask is not None else None),
779
780
                    )
                else:
781
782
783
784
785
786
                    layer_outputs = encoder_layer(
                        hidden_states,
                        attention_mask,
                        layer_head_mask=(head_mask[idx] if head_mask is not None else None),
                        output_attentions=output_attentions,
                    )
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807

                hidden_states = layer_outputs[0]

            if output_attentions:
                all_attentions = all_attentions + (layer_outputs[1],)

        # add final layer norm
        hidden_states = self.layer_norm(hidden_states)

        if output_hidden_states:
            encoder_states = encoder_states + (hidden_states,)

        if not return_dict:
            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
        return BaseModelOutput(
            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
        )


class BlenderbotDecoder(BlenderbotPreTrainedModel):
    """
808
    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`BlenderbotDecoderLayer`]
809
810
811

    Args:
        config: BlenderbotConfig
812
        embed_tokens (nn.Embedding): output embedding
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
    """

    def __init__(self, config: BlenderbotConfig, embed_tokens: Optional[nn.Embedding] = None):
        super().__init__(config)
        self.dropout = config.dropout
        self.layerdrop = config.decoder_layerdrop
        self.padding_idx = config.pad_token_id
        self.max_target_positions = config.max_position_embeddings
        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0

        if embed_tokens is not None:
            self.embed_tokens = embed_tokens
        else:
            self.embed_tokens = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx)

        self.embed_positions = BlenderbotLearnedPositionalEmbedding(
            config.max_position_embeddings,
            config.d_model,
        )
        self.layers = nn.ModuleList([BlenderbotDecoderLayer(config) for _ in range(config.decoder_layers)])
        self.layer_norm = nn.LayerNorm(config.d_model)

835
        self.gradient_checkpointing = False
836
837
        # Initialize weights and apply final processing
        self.post_init()
838

839
840
841
842
843
844
845
846
847
848
849
850
851
852
    def get_input_embeddings(self):
        return self.embed_tokens

    def set_input_embeddings(self, value):
        self.embed_tokens = value

    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, inputs_embeds, past_key_values_length):
        # create causal mask
        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
        combined_attention_mask = None
        if input_shape[-1] > 1:
            combined_attention_mask = _make_causal_mask(
                input_shape, inputs_embeds.dtype, past_key_values_length=past_key_values_length
853
            ).to(inputs_embeds.device)
854
855
856

        if attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
857
858
859
            expanded_attn_mask = _expand_mask(attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]).to(
                inputs_embeds.device
            )
860
861
862
863
864
865
            combined_attention_mask = (
                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
            )

        return combined_attention_mask

866
867
868
869
870
871
    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        encoder_hidden_states=None,
        encoder_attention_mask=None,
872
        head_mask=None,
873
        cross_attn_head_mask=None,
874
875
876
877
878
879
880
881
882
        past_key_values=None,
        inputs_embeds=None,
        use_cache=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        r"""
        Args:
883
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
884
885
886
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

Sylvain Gugger's avatar
Sylvain Gugger committed
887
888
                Indices can be obtained using [`BlenderbotTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.
889

890
891
892
                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
893
894
895
896

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

897
898
                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
899
900
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                of the decoder.
901
            encoder_attention_mask (`torch.LongTensor` of shape `(batch_size, encoder_sequence_length)`, *optional*):
902
                Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. Mask values
903
                selected in `[0, 1]`:
904
905
906
907

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

908
909
                [What are attention masks?](../glossary#attention-mask)
            head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
910
911
                Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0,
                1]`:
912
913

                - 1 indicates the head is **not masked**,
914
                - 0 indicates the head is **masked**.
915

916
            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
917
                Mask to nullify selected heads of the cross-attention modules in the decoder to avoid performing
918
                cross-attention on hidden heads. Mask values selected in `[0, 1]`:
919
920

                - 1 indicates the head is **not masked**,
921
                - 0 indicates the head is **masked**.
922

923
            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Sylvain Gugger's avatar
Sylvain Gugger committed
924
925
926
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
927
928

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
Sylvain Gugger's avatar
Sylvain Gugger committed
929
930
931
932
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.

                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
933
934
                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
Sylvain Gugger's avatar
Sylvain Gugger committed
935
936
937
                `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.
938
939
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
940
                returned tensors for more detail.
941
942
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
943
                for more detail.
944
            return_dict (`bool`, *optional*):
945
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
        """
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # retrieve input_ids and inputs_embeds
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            input_ids = input_ids.view(-1, input_shape[-1])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")

        # 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 inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

971
972
973
        attention_mask = self._prepare_decoder_attention_mask(
            attention_mask, input_shape, inputs_embeds, past_key_values_length
        )
974
975
976
977
978
979
980
981
982
983
984

        # expand encoder attention mask
        if encoder_hidden_states is not None and encoder_attention_mask is not None:
            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
            encoder_attention_mask = _expand_mask(encoder_attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1])

        # embed positions
        positions = self.embed_positions(input_shape, past_key_values_length)

        hidden_states = inputs_embeds + positions

985
        hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
986
987
988
989

        # decoder layers
        all_hidden_states = () if output_hidden_states else None
        all_self_attns = () if output_attentions else None
990
        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
991
        next_decoder_cache = () if use_cache else None
992

993
994
995
        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
        for attn_mask, mask_name in zip([head_mask, cross_attn_head_mask], ["head_mask", "cross_attn_head_mask"]):
            if attn_mask is not None:
996
997
                if attn_mask.size()[0] != len(self.layers):
                    raise ValueError(
Sylvain Gugger's avatar
Sylvain Gugger committed
998
999
                        f"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for"
                        f" {head_mask.size()[0]}."
1000
                    )
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
        for idx, decoder_layer in enumerate(self.layers):
            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
            if output_hidden_states:
                all_hidden_states += (hidden_states,)
            dropout_probability = random.uniform(0, 1)
            if self.training and (dropout_probability < self.layerdrop):
                continue

            past_key_value = past_key_values[idx] if past_key_values is not None else None

1011
            if self.gradient_checkpointing and self.training:
1012

1013
                if use_cache:
1014
                    logger.warning(
1015
                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
1016
                    )
1017
                    use_cache = False
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028

                def create_custom_forward(module):
                    def custom_forward(*inputs):
                        # None for past_key_value
                        return module(*inputs, output_attentions, use_cache)

                    return custom_forward

                layer_outputs = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(decoder_layer),
                    hidden_states,
1029
                    attention_mask,
1030
1031
                    encoder_hidden_states,
                    encoder_attention_mask,
1032
                    head_mask[idx] if head_mask is not None else None,
1033
                    cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None,
1034
1035
1036
1037
1038
1039
                    None,
                )
            else:

                layer_outputs = decoder_layer(
                    hidden_states,
1040
                    attention_mask=attention_mask,
1041
1042
                    encoder_hidden_states=encoder_hidden_states,
                    encoder_attention_mask=encoder_attention_mask,
1043
                    layer_head_mask=(head_mask[idx] if head_mask is not None else None),
1044
1045
1046
                    cross_attn_layer_head_mask=(
                        cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None
                    ),
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
                    past_key_value=past_key_value,
                    output_attentions=output_attentions,
                    use_cache=use_cache,
                )
            hidden_states = layer_outputs[0]

            if use_cache:
                next_decoder_cache += (layer_outputs[3 if output_attentions else 1],)

            if output_attentions:
                all_self_attns += (layer_outputs[1],)
1058
1059
1060

                if encoder_hidden_states is not None:
                    all_cross_attentions += (layer_outputs[2],)
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082

        # add final layer norm
        hidden_states = self.layer_norm(hidden_states)

        # add hidden states from the last decoder layer
        if output_hidden_states:
            all_hidden_states += (hidden_states,)

        next_cache = next_decoder_cache if use_cache else None
        if not return_dict:
            return tuple(
                v
                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
                if v is not None
            )
        return BaseModelOutputWithPastAndCrossAttentions(
            last_hidden_state=hidden_states,
            past_key_values=next_cache,
            hidden_states=all_hidden_states,
            attentions=all_self_attns,
            cross_attentions=all_cross_attentions,
        )
Sam Shleifer's avatar
Sam Shleifer committed
1083
1084
1085


@add_start_docstrings(
1086
1087
    "The bare Blenderbot Model outputting raw hidden-states without any specific head on top.",
    BLENDERBOT_START_DOCSTRING,
1088
)
1089
1090
1091
class BlenderbotModel(BlenderbotPreTrainedModel):
    def __init__(self, config: BlenderbotConfig):
        super().__init__(config)
1092

1093
1094
1095
1096
1097
1098
        padding_idx, vocab_size = config.pad_token_id, config.vocab_size
        self.shared = nn.Embedding(vocab_size, config.d_model, padding_idx)

        self.encoder = BlenderbotEncoder(config, self.shared)
        self.decoder = BlenderbotDecoder(config, self.shared)

1099
1100
        # Initialize weights and apply final processing
        self.post_init()
1101
1102
1103
1104
1105

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
            warnings.warn(
Sylvain Gugger's avatar
Sylvain Gugger committed
1106
1107
1108
                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
                " checkpoint `facebook/small_blenderbot-90M` with"
                " `BlenderbotSmallModel.from_pretrained('facebook/small_blenderbot-90M')` instead.",
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
                FutureWarning,
            )
            return BlenderbotSmallModel.from_pretrained(pretrained_model_name_or_path)

        return super(BlenderbotModel, cls).from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)

    def get_input_embeddings(self):
        return self.shared

    def set_input_embeddings(self, value):
        self.shared = value
        self.encoder.embed_tokens = self.shared
        self.decoder.embed_tokens = self.shared

    def get_encoder(self):
        return self.encoder

    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Union[Tuple, BaseModelOutput]] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
1149
1150
1151
        r"""
        Returns:

1152
        Example:
1153

1154
1155
        ```python
        >>> from transformers import BlenderbotTokenizer, BlenderbotModel
1156

1157
1158
        >>> model = BlenderbotModel.from_pretrained("facebook/blenderbot-400M-distill")
        >>> tokenizer = BlenderbotTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
1159

1160
        >>> inputs = tokenizer("Studies have been shown that owning a dog is good for you", return_tensors="pt")
1161
        >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
1162
        >>> outputs = model(input_ids=inputs.input_ids, decoder_input_ids=decoder_input_ids)
1163

1164
        >>> last_hidden_states = outputs.last_hidden_state
1165
1166
        >>> list(last_hidden_states.shape)
        [1, 6, 1280]
1167
        ```"""
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if encoder_outputs is None:
            encoder_outputs = self.encoder(
                input_ids=input_ids,
                attention_mask=attention_mask,
1179
                head_mask=head_mask,
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
                inputs_embeds=inputs_embeds,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict=return_dict,
            )
        # If the user passed a tuple for encoder_outputs, we wrap it in a BaseModelOutput when return_dict=True
        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
            encoder_outputs = BaseModelOutput(
                last_hidden_state=encoder_outputs[0],
                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
            )

        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            encoder_hidden_states=encoder_outputs[0],
            encoder_attention_mask=attention_mask,
1199
            head_mask=decoder_head_mask,
1200
            cross_attn_head_mask=cross_attn_head_mask,
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
            past_key_values=past_key_values,
            inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        if not return_dict:
            return decoder_outputs + encoder_outputs

        return Seq2SeqModelOutput(
            last_hidden_state=decoder_outputs.last_hidden_state,
            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,
            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
            encoder_hidden_states=encoder_outputs.hidden_states,
            encoder_attentions=encoder_outputs.attentions,
        )
1222
1223
1224


@add_start_docstrings(
1225
    "The Blenderbot Model with a language modeling head. Can be used for summarization.", BLENDERBOT_START_DOCSTRING
Sam Shleifer's avatar
Sam Shleifer committed
1226
)
1227
1228
1229
1230
class BlenderbotForConditionalGeneration(BlenderbotPreTrainedModel):
    base_model_prefix = "model"
    _keys_to_ignore_on_load_missing = [
        r"final_logits_bias",
1231
1232
1233
        r"encoder.version",
        r"decoder.version",
        r"lm_head.weight",
1234
    ]
Sam Shleifer's avatar
Sam Shleifer committed
1235

1236
1237
1238
1239
1240
1241
    def __init__(self, config: BlenderbotConfig):
        super().__init__(config)
        self.model = BlenderbotModel(config)
        self.register_buffer("final_logits_bias", torch.zeros((1, self.model.shared.num_embeddings)))
        self.lm_head = nn.Linear(config.d_model, self.model.shared.num_embeddings, bias=False)

1242
1243
        # Initialize weights and apply final processing
        self.post_init()
1244
1245
1246
1247
1248

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], *model_args, **kwargs):
        if pretrained_model_name_or_path == "facebook/blenderbot-90M":
            warnings.warn(
Sylvain Gugger's avatar
Sylvain Gugger committed
1249
1250
1251
                "The checkpoint `facebook/blenderbot-90M` is deprecated. In the future, please use the identical"
                " checkpoint `facebook/small_blenderbot-90M` with"
                " `BlenderbotSmallForConditionalGeneration.from_pretrained('facebook/small_blenderbot-90M')` instead.",
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
                FutureWarning,
            )
            return BlenderbotSmallForConditionalGeneration.from_pretrained(pretrained_model_name_or_path)

        return super(BlenderbotForConditionalGeneration, cls).from_pretrained(
            pretrained_model_name_or_path, *model_args, **kwargs
        )

    def get_encoder(self):
        return self.model.get_encoder()

    def get_decoder(self):
        return self.model.get_decoder()

    def resize_token_embeddings(self, new_num_tokens: int) -> nn.Embedding:
        new_embeddings = super().resize_token_embeddings(new_num_tokens)
        self._resize_final_logits_bias(new_num_tokens)
        return new_embeddings

    def _resize_final_logits_bias(self, new_num_tokens: int) -> None:
        old_num_tokens = self.final_logits_bias.shape[-1]
        if new_num_tokens <= old_num_tokens:
            new_bias = self.final_logits_bias[:, :new_num_tokens]
        else:
            extra_bias = torch.zeros((1, new_num_tokens - old_num_tokens), device=self.final_logits_bias.device)
            new_bias = torch.cat([self.final_logits_bias, extra_bias], dim=1)
        self.register_buffer("final_logits_bias", new_bias)

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    @add_start_docstrings_to_model_forward(BLENDERBOT_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    @add_end_docstrings(BLENDERBOT_GENERATION_EXAMPLE)
    def forward(
        self,
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[Union[Tuple, BaseModelOutput]] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
1308
        r"""
1309
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1310
1311
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
1312
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
1313
1314
1315
1316
1317
1318

        Returns:
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if labels is not None:
1319
1320
1321
            if use_cache:
                logger.warning("The `use_cache` argument is changed to `False` since `labels` is provided.")
            use_cache = False
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
            if decoder_input_ids is None:
                decoder_input_ids = shift_tokens_right(
                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
                )

        outputs = self.model(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            encoder_outputs=encoder_outputs,
            decoder_attention_mask=decoder_attention_mask,
1333
1334
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
1335
            cross_attn_head_mask=cross_attn_head_mask,
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        lm_logits = self.lm_head(outputs[0]) + self.final_logits_bias

        masked_lm_loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            masked_lm_loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (lm_logits,) + outputs[1:]
            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output

        return Seq2SeqLMOutput(
            loss=masked_lm_loss,
            logits=lm_logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )

    def prepare_inputs_for_generation(
1368
1369
1370
1371
1372
        self,
        decoder_input_ids,
        past=None,
        attention_mask=None,
        head_mask=None,
1373
1374
        decoder_head_mask=None,
        cross_attn_head_mask=None,
1375
1376
1377
        use_cache=None,
        encoder_outputs=None,
        **kwargs
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
    ):
        # cut decoder_input_ids if past is used
        if past is not None:
            decoder_input_ids = decoder_input_ids[:, -1:]

        return {
            "input_ids": None,  # encoder_outputs is defined. input_ids not needed
            "encoder_outputs": encoder_outputs,
            "past_key_values": past,
            "decoder_input_ids": decoder_input_ids,
            "attention_mask": attention_mask,
1389
            "head_mask": head_mask,
1390
1391
            "decoder_head_mask": decoder_head_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
1392
1393
            "use_cache": use_cache,  # change this to avoid caching (presumably for debugging)
        }
Sam Shleifer's avatar
Sam Shleifer committed
1394

1395
1396
1397
1398
1399
1400
1401
1402
1403
    @staticmethod
    def _reorder_cache(past, beam_idx):
        reordered_past = ()
        for layer_past in past:
            # cached cross_attention states don't have to be reordered -> they are always the same
            reordered_past += (
                tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
            )
        return reordered_past
1404
1405
1406
1407
1408
1409


# Copied from transformers.models.bart.modeling_bart.BartDecoderWrapper with Bart->Blenderbot
class BlenderbotDecoderWrapper(BlenderbotPreTrainedModel):
    """
    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
1410
    used in combination with the [`EncoderDecoderModel`] framework.
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
    """

    def __init__(self, config):
        super().__init__(config)
        self.decoder = BlenderbotDecoder(config)

    def forward(self, *args, **kwargs):
        return self.decoder(*args, **kwargs)


1421
# Copied from transformers.models.bart.modeling_bart.BartForCausalLM with Bart->Blenderbot, facebook/bart-base->facebook/blenderbot-400M-distill
1422
1423
1424
1425
1426
class BlenderbotForCausalLM(BlenderbotPreTrainedModel):
    def __init__(self, config):
        config = copy.deepcopy(config)
        config.is_decoder = True
        config.is_encoder_decoder = False
1427
        super().__init__(config)
1428
1429
1430
1431
        self.model = BlenderbotDecoderWrapper(config)

        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)

1432
1433
        # Initialize weights and apply final processing
        self.post_init()
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455

    def get_input_embeddings(self):
        return self.model.decoder.embed_tokens

    def set_input_embeddings(self, value):
        self.model.decoder.embed_tokens = value

    def get_output_embeddings(self):
        return self.lm_head

    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    def set_decoder(self, decoder):
        self.model.decoder = decoder

    def get_decoder(self):
        return self.model.decoder

    @replace_return_docstrings(output_type=CausalLMOutputWithCrossAttentions, config_class=_CONFIG_FOR_DOC)
    def forward(
        self,
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithCrossAttentions]:
1470
1471
        r"""
        Args:
1472
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
1473
1474
1475
                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
                provide it.

Sylvain Gugger's avatar
Sylvain Gugger committed
1476
1477
                Indices can be obtained using [`BlenderbotTokenizer`]. See [`PreTrainedTokenizer.encode`] and
                [`PreTrainedTokenizer.__call__`] for details.
1478

1479
1480
1481
                [What are input IDs?](../glossary#input-ids)
            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
1482
1483
1484
1485

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.

1486
1487
                [What are attention masks?](../glossary#attention-mask)
            encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
1488
1489
                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
                if the model is configured as a decoder.
1490
            encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
1491
                Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
1492
1493
1494
                in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
            head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
1495
1496

                - 1 indicates the head is **not masked**,
1497
                - 0 indicates the head is **masked**.
1498

1499
1500
            cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
                Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
1501
1502

                - 1 indicates the head is **not masked**,
1503
                - 0 indicates the head is **masked**.
1504

1505
            past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
Sylvain Gugger's avatar
Sylvain Gugger committed
1506
1507
1508
1509
                Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
                shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of
                shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional
                tensors are only required when the model is used as a decoder in a Sequence to Sequence model.
1510
1511

                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
Sylvain Gugger's avatar
Sylvain Gugger committed
1512
                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
1513

Sylvain Gugger's avatar
Sylvain Gugger committed
1514
1515
1516
                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)`.
1517
            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1518
1519
1520
                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
                config.vocab_size]` or -100 (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]`.
1521
            use_cache (`bool`, *optional*):
Sylvain Gugger's avatar
Sylvain Gugger committed
1522
1523
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
1524
1525
1526

                - 1 for tokens that are **not masked**,
                - 0 for tokens that are **masked**.
1527
1528
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
1529
                returned tensors for more detail.
1530
1531
            output_hidden_states (`bool`, *optional*):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
1532
                for more detail.
1533
            return_dict (`bool`, *optional*):
1534
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
1535
1536
1537

        Returns:

1538
        Example:
1539

1540
1541
        ```python
        >>> from transformers import BlenderbotTokenizer, BlenderbotForCausalLM
1542

1543
1544
1545
1546
        >>> tokenizer = BlenderbotTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
        >>> model = BlenderbotForCausalLM.from_pretrained(
        ...     "facebook/blenderbot-400M-distill", add_cross_attention=False
        ... )
1547
1548
1549
        >>> assert model.config.is_decoder, f"{model.__class__} has to be configured as a decoder."
        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)
1550

1551
        >>> logits = outputs.logits
1552
1553
1554
        >>> expected_shape = [1, inputs.input_ids.shape[-1], model.config.vocab_size]
        >>> list(logits.shape) == expected_shape
        True
1555
        ```"""
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569

        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model.decoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            encoder_hidden_states=encoder_hidden_states,
            encoder_attention_mask=encoder_attention_mask,
            head_mask=head_mask,
1570
            cross_attn_head_mask=cross_attn_head_mask,
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        logits = self.lm_head(outputs[0])

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.config.vocab_size), labels.view(-1))

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithCrossAttentions(
            loss=loss,
            logits=logits,
            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, past=None, attention_mask=None, use_cache=None, **kwargs):
        # 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_ids.shape)

        if past:
            input_ids = input_ids[:, -1:]
        # first step, decoder_cached_states are empty
        return {
            "input_ids": input_ids,  # encoder_outputs is defined. input_ids not needed
            "attention_mask": attention_mask,
            "past_key_values": past,
            "use_cache": use_cache,
        }

    @staticmethod
    def _reorder_cache(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