utils.py 251 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
# coding=utf-8
# Copyright 2020 The Google AI Language Team Authors, Facebook AI Research authors and The HuggingFace Inc. team.
# Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# 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.

17
import copy
18
19
20
import inspect
import warnings
from dataclasses import dataclass
21
from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
22
23
24
25
26

import torch
import torch.distributed as dist
from torch import nn

27
from ..deepspeed import is_deepspeed_zero3_enabled
28
29
30
31
32
33
34
35
36
from ..modeling_outputs import CausalLMOutputWithPast, Seq2SeqLMOutput
from ..models.auto import (
    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
    MODEL_FOR_CAUSAL_LM_MAPPING,
    MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
    MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
    MODEL_FOR_VISION_2_SEQ_MAPPING,
)
from ..utils import ModelOutput, logging
37
from .beam_constraints import DisjunctiveConstraint, PhrasalConstraint
38
from .beam_search import BeamScorer, BeamSearchScorer, ConstrainedBeamSearchScorer
39
from .configuration_utils import GenerationConfig
40
from .logits_process import (
Sanchit Gandhi's avatar
Sanchit Gandhi committed
41
    ClassifierFreeGuidanceLogitsProcessor,
42
    EncoderNoRepeatNGramLogitsProcessor,
Karim Foda's avatar
Karim Foda committed
43
    EncoderRepetitionPenaltyLogitsProcessor,
44
45
    EpsilonLogitsWarper,
    EtaLogitsWarper,
46
47
48
49
50
51
52
53
54
    ExponentialDecayLengthPenalty,
    ForcedBOSTokenLogitsProcessor,
    ForcedEOSTokenLogitsProcessor,
    ForceTokensLogitsProcessor,
    HammingDiversityLogitsProcessor,
    InfNanRemoveLogitsProcessor,
    LogitNormalization,
    LogitsProcessorList,
    MinLengthLogitsProcessor,
55
    MinNewTokensLengthLogitsProcessor,
56
57
58
59
    NoBadWordsLogitsProcessor,
    NoRepeatNGramLogitsProcessor,
    PrefixConstrainedLogitsProcessor,
    RepetitionPenaltyLogitsProcessor,
60
    SequenceBiasLogitsProcessor,
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
    SuppressTokensAtBeginLogitsProcessor,
    SuppressTokensLogitsProcessor,
    TemperatureLogitsWarper,
    TopKLogitsWarper,
    TopPLogitsWarper,
    TypicalLogitsWarper,
)
from .stopping_criteria import (
    MaxLengthCriteria,
    MaxTimeCriteria,
    StoppingCriteria,
    StoppingCriteriaList,
    validate_stopping_criteria,
)


77
if TYPE_CHECKING:
78
    from ..modeling_utils import PreTrainedModel
79
80
    from .streamers import BaseStreamer

81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
logger = logging.get_logger(__name__)


@dataclass
class GreedySearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using greedy search.


    Args:
        sequences (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size, config.vocab_size)`.
        attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class ContrastiveSearchEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using contrastive search.

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size, config.vocab_size)`.
        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads,
            sequence_length, sequence_length)`.
        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.
        decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class ContrastiveSearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using contrastive search.

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when
        `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size, config.vocab_size)`.
        attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is
        passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class GreedySearchEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of encoder-decoder generation models using greedy search. Hidden states and attention
    weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the
    encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)


    Args:
        sequences (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size, config.vocab_size)`.
        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads,
            sequence_length, sequence_length)`.
        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size, sequence_length, hidden_size)`.
        decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class SampleDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using sampling.


    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size*num_return_sequences, config.vocab_size)`.
        attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(num_return_sequences*batch_size, num_heads, generated_length,
            sequence_length)`.
        hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(num_return_sequences*batch_size, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class SampleEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of encoder-decoder generation models using sampling. Hidden states and attention weights of
    the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states
    attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)


    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax)
            at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for
            each generated token), with each tensor of shape `(batch_size*num_return_sequences, config.vocab_size)`.
        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape
            `(batch_size*num_return_sequences, num_heads, sequence_length, sequence_length)`.
        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size*num_return_sequences, sequence_length, hidden_size)`.
        decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_return_sequences, num_heads, generated_length,
            sequence_length)`.
        cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_return_sequences, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSearchDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using beam search.

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        sequences_scores (`torch.FloatTensor` of shape `(batch_size*num_return_sequences)`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Final beam scores of the generated `sequences`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting
            of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam.
            Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token),
            with each tensor of shape `(batch_size*num_beams*num_return_sequences, config.vocab_size)`.
310
        beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
311
            Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
312
            `(batch_size*num_return_sequences, sequence_length)`.
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
        attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams, num_heads, generated_length, sequence_length)`.
        hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSearchEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of encoder-decoder generation models using beam search. Hidden states and attention weights
    of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the encoder_hidden_states
    attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        sequences_scores (`torch.FloatTensor` of shape `(batch_size*num_return_sequences)`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Final beam scores of the generated `sequences`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting
            of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam.
            Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token),
            with each tensor of shape `(batch_size*num_beams, config.vocab_size)`.
347
        beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
348
            Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
349
            `(batch_size*num_return_sequences, sequence_length)`.
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
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads,
            sequence_length, sequence_length)`.
        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size*num_beams*num_return_sequences, sequence_length, hidden_size)`.
        decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, num_heads, generated_length,
            sequence_length)`.
        cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSampleDecoderOnlyOutput(ModelOutput):
    """
    Base class for outputs of decoder-only generation models using beam sample.

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        sequences_scores (`torch.FloatTensor` of shape `(batch_size * num_return_sequence)`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Final beam scores of the generated `sequences`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting
            of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam.
            Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token),
            with each tensor of shape `(batch_size*num_beams*num_return_sequences, config.vocab_size)`.
395
        beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
396
            Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
397
            `(batch_size*num_return_sequences, sequence_length)`.
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
        attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams, num_heads, generated_length, sequence_length)`.
        hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


@dataclass
class BeamSampleEncoderDecoderOutput(ModelOutput):
    """
    Base class for outputs of encoder-decoder generation models using beam sampling. Hidden states and attention
    weights of the decoder (respectively the encoder) can be accessed via the encoder_attentions and the
    encoder_hidden_states attributes (respectively the decoder_attentions and the decoder_hidden_states attributes)

    Args:
        sequences (`torch.LongTensor` of shape `(batch_size*num_beams, sequence_length)`):
            The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter
            if all batches finished early due to the `eos_token_id`.
        sequences_scores (`torch.FloatTensor` of shape `(batch_size * num_return_sequence)`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Final beam scores of the generated `sequences`.
        scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting
            of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam.
            Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token),
            with each tensor of shape `(batch_size*num_beams, config.vocab_size)`).
        beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True` is passed or when `config.output_scores=True`):
            Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
434
            `(batch_size*num_return_sequences, sequence_length)`.
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
        encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads,
            sequence_length, sequence_length)`.
        encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
            shape `(batch_size*num_beams, sequence_length, hidden_size)`.
        decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams, num_heads, generated_length, sequence_length)`.
        cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True` is passed or `config.output_attentions=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`.
        decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
            Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of
            `torch.FloatTensor` of shape `(batch_size*num_beams, generated_length, hidden_size)`.
    """

    sequences: torch.LongTensor = None
    sequences_scores: Optional[torch.FloatTensor] = None
    scores: Optional[Tuple[torch.FloatTensor]] = None
    beam_indices: Optional[torch.LongTensor] = None
    encoder_attentions: Optional[Tuple[torch.FloatTensor]] = None
    encoder_hidden_states: Optional[Tuple[torch.FloatTensor]] = None
    decoder_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    cross_attentions: Optional[Tuple[Tuple[torch.FloatTensor]]] = None
    decoder_hidden_states: Optional[Tuple[Tuple[torch.FloatTensor]]] = None


GreedySearchOutput = Union[GreedySearchEncoderDecoderOutput, GreedySearchDecoderOnlyOutput]
SampleOutput = Union[SampleEncoderDecoderOutput, SampleDecoderOnlyOutput]
BeamSearchOutput = Union[BeamSearchEncoderDecoderOutput, BeamSearchDecoderOnlyOutput]
BeamSampleOutput = Union[BeamSampleEncoderDecoderOutput, BeamSampleDecoderOnlyOutput]
ContrastiveSearchOutput = Union[ContrastiveSearchEncoderDecoderOutput, ContrastiveSearchDecoderOnlyOutput]
GenerateOutput = Union[GreedySearchOutput, SampleOutput, BeamSearchOutput, BeamSampleOutput, ContrastiveSearchOutput]


class GenerationMixin:
    """
    A class containing all functions for auto-regressive text generation, to be used as a mixin in [`PreTrainedModel`].

    The class exposes [`~generation.GenerationMixin.generate`], which can be used for:
        - *greedy decoding* by calling [`~generation.GenerationMixin.greedy_search`] if `num_beams=1` and
477
          `do_sample=False`
478
479
480
        - *contrastive search* by calling [`~generation.GenerationMixin.contrastive_search`] if `penalty_alpha>0` and
          `top_k>1`
        - *multinomial sampling* by calling [`~generation.GenerationMixin.sample`] if `num_beams=1` and
481
          `do_sample=True`
482
        - *beam-search decoding* by calling [`~generation.GenerationMixin.beam_search`] if `num_beams>1` and
483
          `do_sample=False`
484
        - *beam-search multinomial sampling* by calling [`~generation.GenerationMixin.beam_sample`] if `num_beams>1`
485
          and `do_sample=True`
486
        - *diverse beam-search decoding* by calling [`~generation.GenerationMixin.group_beam_search`], if `num_beams>1`
487
          and `num_beam_groups>1`
488
        - *constrained beam-search decoding* by calling [`~generation.GenerationMixin.constrained_beam_search`], if
489
490
491
          `constraints!=None` or `force_words_ids!=None`

    You do not need to call any of the above methods directly. Pass custom parameter values to 'generate' instead. To
492
    learn more about decoding strategies refer to the [text generation strategies guide](../generation_strategies).
493
494
    """

495
496
    def prepare_inputs_for_generation(self, *args, **kwargs):
        raise NotImplementedError(
497
            "A model class needs to define a `prepare_inputs_for_generation` method in order to use `.generate()`."
498
499
        )

500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
    def _prepare_model_inputs(
        self,
        inputs: Optional[torch.Tensor] = None,
        bos_token_id: Optional[int] = None,
        model_kwargs: Optional[Dict[str, torch.Tensor]] = None,
    ) -> Tuple[torch.Tensor, Optional[str], Dict[str, torch.Tensor]]:
        """
        This function extracts the model-specific `inputs` for generation.
        """
        # 1. retrieve all kwargs that are non-None or non-model input related.
        # some encoder-decoder models have different names for model and encoder
        if (
            self.config.is_encoder_decoder
            and hasattr(self, "encoder")
            and self.encoder.main_input_name != self.main_input_name
        ):
            input_name = self.encoder.main_input_name
        else:
            input_name = self.main_input_name

        model_kwargs = {k: v for k, v in model_kwargs.items() if v is not None or k != input_name}

        # 2. check whether model_input_name is passed as kwarg
        # if yes and `inputs` is None use kwarg inputs
        inputs_kwarg = model_kwargs.pop(input_name, None)
        if inputs_kwarg is not None and inputs is not None:
            raise ValueError(
527
                f"`inputs`: {inputs}` were passed alongside {input_name} which is not allowed."
528
529
530
531
532
                f"Make sure to either pass {inputs} or {input_name}=..."
            )
        elif inputs_kwarg is not None:
            inputs = inputs_kwarg

533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
        # 3. In the presence of `inputs_embeds` for text models:
        # - decoder-only models should complain if the user attempts to pass `inputs_embeds`, but the model
        # doesn't have its forwarding implemented. `inputs_embeds` is kept in `model_kwargs` and can coexist with
        # input_ids (`inputs_embeds` will be used in the 1st generation step, as opposed to `input_ids`)
        # - encoder-decoder models should complain if the user attempts to pass `inputs_embeds` and `input_ids`, and
        # pull the former to inputs. It will be used in place of `input_ids` to get the encoder hidden states.
        if input_name == "input_ids" and "inputs_embeds" in model_kwargs:
            if not self.config.is_encoder_decoder:
                has_inputs_embeds_forwarding = "inputs_embeds" in set(
                    inspect.signature(self.prepare_inputs_for_generation).parameters.keys()
                )
                if not has_inputs_embeds_forwarding:
                    raise ValueError(
                        f"You passed `inputs_embeds` to `.generate()`, but the model class {self.__class__.__name__} "
                        "doesn't have its forwarding implemented. See the GPT2 implementation for an example "
                        "(https://github.com/huggingface/transformers/pull/21405), and feel free to open a PR with it!"
                    )
550
551
552
                # In this case, `input_ids` is moved to the `model_kwargs`, so a few automations (like the creation of
                # the attention mask) can rely on the actual model input.
                model_kwargs["input_ids"] = self._maybe_initialize_input_ids_for_generation(
553
                    inputs, bos_token_id, model_kwargs=model_kwargs
554
                )
555
556
557
            else:
                if inputs is not None:
                    raise ValueError("You passed `inputs_embeds` and `input_ids` to `.generate()`. Please pick one.")
558
            inputs, input_name = model_kwargs["inputs_embeds"], "inputs_embeds"
559
560

        # 4. if `inputs` is still None, try to create `input_ids` from BOS token
561
        inputs = self._maybe_initialize_input_ids_for_generation(inputs, bos_token_id, model_kwargs)
562
563
564
565
566
567
568
569
        return inputs, input_name, model_kwargs

    def adjust_logits_during_generation(self, logits: torch.FloatTensor, **kwargs) -> torch.FloatTensor:
        """
        Implement in subclasses of [`PreTrainedModel`] for custom behavior to adjust the logits in the generate method.
        """
        return logits

570
571
572
573
    def _maybe_initialize_input_ids_for_generation(
        self,
        inputs: Optional[torch.Tensor] = None,
        bos_token_id: Optional[int] = None,
574
        model_kwargs: Optional[Dict[str, torch.Tensor]] = None,
575
    ) -> torch.LongTensor:
576
577
578
579
        """Initializes input ids for generation, if necessary."""
        if inputs is not None:
            return inputs

580
        encoder_outputs = model_kwargs.get("encoder_outputs")
581
582
583
584
585
586
587
        if self.config.is_encoder_decoder and encoder_outputs is not None:
            # make dummy input_ids with value -100, as a sanity check ensuring that they won't be used for encoding
            shape = encoder_outputs.last_hidden_state.size()[:-1]
            return torch.ones(shape, dtype=torch.long, device=self.device) * -100

        if bos_token_id is None:
            raise ValueError("`bos_token_id` has to be defined when no `input_ids` are provided.")
588

589
590
591
592
593
594
595
        # If there is some tensor in `model_kwargs`, we can infer the batch size from it. This is helpful with
        # soft-prompting or in multimodal implementations built on top of decoder-only language models.
        batch_size = 1
        for value in model_kwargs.values():
            if isinstance(value, torch.Tensor):
                batch_size = value.shape[0]
                break
596
        return torch.ones((batch_size, 1), dtype=torch.long, device=self.device) * bos_token_id
597
598
599
600
601

    def _prepare_attention_mask_for_generation(
        self,
        inputs: torch.Tensor,
        pad_token_id: Optional[int],
602
        eos_token_id: Optional[Union[int, List[int]]],
603
604
605
    ) -> torch.LongTensor:
        is_input_ids = len(inputs.shape) == 2 and inputs.dtype in [torch.int, torch.long]
        is_pad_token_in_inputs = (pad_token_id is not None) and (pad_token_id in inputs)
606
607
608
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        is_pad_token_not_equal_to_eos_token_id = (eos_token_id is None) or (pad_token_id not in eos_token_id)
609
610
611
612
613
614
615
616
617
618
619
620

        # Check if input is input_ids and padded -> only then is attention_mask defined
        if is_input_ids and is_pad_token_in_inputs and is_pad_token_not_equal_to_eos_token_id:
            return inputs.ne(pad_token_id).long()
        else:
            return torch.ones(inputs.shape[:2], dtype=torch.long, device=inputs.device)

    def _prepare_encoder_decoder_kwargs_for_generation(
        self, inputs_tensor: torch.Tensor, model_kwargs, model_input_name: Optional[str] = None
    ) -> Dict[str, Any]:
        # 1. get encoder
        encoder = self.get_encoder()
621
622
623
624
        # Compatibility with Accelerate big model inference: we need the encoder to outputs stuff on the same device
        # as the inputs.
        if hasattr(encoder, "_hf_hook"):
            encoder._hf_hook.io_same_device = True
625

626
        # 2. Prepare encoder args and encoder kwargs from model kwargs.
627
628
629
630
631
632
        irrelevant_prefix = ["decoder_", "cross_attn", "use_cache"]
        encoder_kwargs = {
            argument: value
            for argument, value in model_kwargs.items()
            if not any(argument.startswith(p) for p in irrelevant_prefix)
        }
633
634
635
636
637
638
        encoder_signature = set(inspect.signature(encoder.forward).parameters)
        encoder_accepts_wildcard = "kwargs" in encoder_signature or "model_kwargs" in encoder_signature
        if not encoder_accepts_wildcard:
            encoder_kwargs = {
                argument: value for argument, value in encoder_kwargs.items() if argument in encoder_signature
            }
639
640
641
642
643
644
645
646
647
648
649
650

        # 3. make sure that encoder returns `ModelOutput`
        model_input_name = model_input_name if model_input_name is not None else self.main_input_name
        encoder_kwargs["return_dict"] = True
        encoder_kwargs[model_input_name] = inputs_tensor
        model_kwargs["encoder_outputs"]: ModelOutput = encoder(**encoder_kwargs)

        return model_kwargs

    def _prepare_decoder_input_ids_for_generation(
        self,
        batch_size: int,
651
652
        model_input_name: str,
        model_kwargs: Dict[str, torch.Tensor],
653
654
655
        decoder_start_token_id: int = None,
        bos_token_id: int = None,
        device: torch.device = None,
656
657
658
659
    ) -> Tuple[torch.LongTensor, Dict[str, torch.Tensor]]:
        """Prepares `decoder_input_ids` for generation with encoder-decoder models"""
        # 1. Check whether the user has defined `decoder_input_ids` manually. To facilitate in terms of input naming,
        # we also allow the user to pass it under `input_ids`, if the encoder does not use it as the main input.
660
        if model_kwargs is not None and "decoder_input_ids" in model_kwargs:
661
662
663
            decoder_input_ids = model_kwargs.pop("decoder_input_ids")
        elif "input_ids" in model_kwargs and model_input_name != "input_ids":
            decoder_input_ids = model_kwargs.pop("input_ids")
664
        else:
665
666
667
668
669
670
671
672
673
674
675
            decoder_input_ids = None

        # 2. Encoder-decoder models expect the `decoder_input_ids` to start with a special token. Let's ensure that.
        decoder_start_token_id = self._get_decoder_start_token_id(decoder_start_token_id, bos_token_id)
        if device is None:
            device = self.device
        decoder_input_ids_start = torch.ones((batch_size, 1), dtype=torch.long, device=device) * decoder_start_token_id

        # no user input -> use decoder_start_token_id as decoder_input_ids
        if decoder_input_ids is None:
            decoder_input_ids = decoder_input_ids_start
676
677
678
        # exception: Donut checkpoints have task-specific decoder starts and don't expect a BOS token
        elif self.config.model_type == "vision-encoder-decoder" and "donut" in self.name_or_path.lower():
            pass
679
680
681
682
683
684
685
686
687
688
689
690
691
        # user input but doesn't start with decoder_start_token_id -> prepend decoder_start_token_id (and adjust
        # decoder_attention_mask if provided)
        elif (decoder_input_ids[:, 0] != decoder_start_token_id).all().item():
            decoder_input_ids = torch.cat([decoder_input_ids_start, decoder_input_ids], dim=-1)
            if "decoder_attention_mask" in model_kwargs:
                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
                decoder_attention_mask = torch.cat(
                    (torch.ones_like(decoder_attention_mask)[:, :1], decoder_attention_mask),
                    dim=-1,
                )
                model_kwargs["decoder_attention_mask"] = decoder_attention_mask

        return decoder_input_ids, model_kwargs
692
693
694

    def _get_decoder_start_token_id(self, decoder_start_token_id: int = None, bos_token_id: int = None) -> int:
        decoder_start_token_id = (
695
696
697
            decoder_start_token_id
            if decoder_start_token_id is not None
            else self.generation_config.decoder_start_token_id
698
        )
699
        bos_token_id = bos_token_id if bos_token_id is not None else self.generation_config.bos_token_id
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716

        if decoder_start_token_id is not None:
            return decoder_start_token_id
        elif bos_token_id is not None:
            return bos_token_id
        raise ValueError(
            "`decoder_start_token_id` or `bos_token_id` has to be defined for encoder-decoder generation."
        )

    @staticmethod
    def _expand_inputs_for_generation(
        expand_size: int = 1,
        is_encoder_decoder: bool = False,
        input_ids: Optional[torch.LongTensor] = None,
        **model_kwargs,
    ) -> Tuple[torch.LongTensor, Dict[str, Any]]:
        """Expands tensors from [batch_size, ...] to [batch_size * expand_size, ...]"""
717
718
719
720
721
722
723

        def _expand_dict_for_generation(dict_to_expand):
            for key in dict_to_expand:
                if dict_to_expand[key] is not None and isinstance(dict_to_expand[key], torch.Tensor):
                    dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
            return dict_to_expand

724
725
726
        if input_ids is not None:
            input_ids = input_ids.repeat_interleave(expand_size, dim=0)

727
        model_kwargs = _expand_dict_for_generation(model_kwargs)
728
729

        if is_encoder_decoder:
730
            if model_kwargs.get("encoder_outputs") is None:
731
                raise ValueError("If `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.")
732
            model_kwargs["encoder_outputs"] = _expand_dict_for_generation(model_kwargs["encoder_outputs"])
733
734
735

        return input_ids, model_kwargs

736
    def _extract_past_from_model_output(self, outputs: ModelOutput, standardize_cache_format: bool = False):
737
        past_key_values = None
738
        if "past_key_values" in outputs:
739
            past_key_values = outputs.past_key_values
740
        elif "mems" in outputs:
741
            past_key_values = outputs.mems
742
        elif "past_buckets_states" in outputs:
743
            past_key_values = outputs.past_buckets_states
744
745
746
747

        # Bloom fix: standardizes the cache format when requested
        if standardize_cache_format and hasattr(self, "_convert_to_standard_cache"):
            batch_size = outputs.logits.shape[0]
748
749
            past_key_values = self._convert_to_standard_cache(past_key_values, batch_size=batch_size)
        return past_key_values
750
751

    def _update_model_kwargs_for_generation(
752
753
754
755
756
        self,
        outputs: ModelOutput,
        model_kwargs: Dict[str, Any],
        is_encoder_decoder: bool = False,
        standardize_cache_format: bool = False,
757
    ) -> Dict[str, Any]:
758
759
        # update past_key_values
        model_kwargs["past_key_values"] = self._extract_past_from_model_output(
760
761
            outputs, standardize_cache_format=standardize_cache_format
        )
Sylvain Gugger's avatar
Sylvain Gugger committed
762
763
        if getattr(outputs, "state", None) is not None:
            model_kwargs["state"] = outputs.state
764
765
766
767
768
769
770

        # update token_type_ids with last value
        if "token_type_ids" in model_kwargs:
            token_type_ids = model_kwargs["token_type_ids"]
            model_kwargs["token_type_ids"] = torch.cat([token_type_ids, token_type_ids[:, -1].unsqueeze(-1)], dim=-1)

        if not is_encoder_decoder:
771
            # update attention mask
772
773
774
775
776
            if "attention_mask" in model_kwargs:
                attention_mask = model_kwargs["attention_mask"]
                model_kwargs["attention_mask"] = torch.cat(
                    [attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
                )
777
778
779
780
781
782
783
784
        else:
            # update decoder attention mask
            if "decoder_attention_mask" in model_kwargs:
                decoder_attention_mask = model_kwargs["decoder_attention_mask"]
                model_kwargs["decoder_attention_mask"] = torch.cat(
                    [decoder_attention_mask, decoder_attention_mask.new_ones((decoder_attention_mask.shape[0], 1))],
                    dim=-1,
                )
785
786
787

        return model_kwargs

788
    def _reorder_cache(self, past_key_values, beam_idx):
789
790
791
792
793
794
795
        raise NotImplementedError(
            f"Make sure that a `_reorder_cache` function is correctly implemented in {self.__class__.__module__} to"
            f" enable beam search for {self.__class__}"
        )

    def _get_logits_warper(
        self,
796
        generation_config: GenerationConfig,
797
798
799
800
801
802
803
804
805
806
807
    ) -> LogitsProcessorList:
        """
        This class returns a [`LogitsProcessorList`] list object that contains all relevant [`LogitsWarper`] instances
        used for multinomial sampling.
        """

        # instantiate warpers list
        warpers = LogitsProcessorList()

        # the following idea is largely copied from this PR: https://github.com/huggingface/transformers/pull/5420/files
        # all samplers can be found in `generation_utils_samplers.py`
808
809
        if generation_config.temperature is not None and generation_config.temperature != 1.0:
            warpers.append(TemperatureLogitsWarper(generation_config.temperature))
810
        min_tokens_to_keep = 2 if generation_config.num_beams > 1 else 1
811
        if generation_config.top_k is not None and generation_config.top_k != 0:
812
813
814
815
            warpers.append(TopKLogitsWarper(top_k=generation_config.top_k, min_tokens_to_keep=min_tokens_to_keep))
        if generation_config.top_p is not None and generation_config.top_p < 1.0:
            warpers.append(TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=min_tokens_to_keep))
        if generation_config.typical_p is not None and generation_config.typical_p < 1.0:
816
            warpers.append(
817
                TypicalLogitsWarper(mass=generation_config.typical_p, min_tokens_to_keep=min_tokens_to_keep)
818
            )
819
        if generation_config.epsilon_cutoff is not None and 0.0 < generation_config.epsilon_cutoff < 1.0:
820
            warpers.append(
821
                EpsilonLogitsWarper(epsilon=generation_config.epsilon_cutoff, min_tokens_to_keep=min_tokens_to_keep)
822
            )
823
        if generation_config.eta_cutoff is not None and 0.0 < generation_config.eta_cutoff < 1.0:
824
            warpers.append(
825
                EtaLogitsWarper(epsilon=generation_config.eta_cutoff, min_tokens_to_keep=min_tokens_to_keep)
826
            )
827
        # `LogitNormalization` should always be the last logit processor, when present
828
        if generation_config.renormalize_logits is True:
829
830
831
832
833
            warpers.append(LogitNormalization())
        return warpers

    def _get_logits_processor(
        self,
834
        generation_config: GenerationConfig,
835
836
837
838
839
840
841
842
843
844
        input_ids_seq_length: int,
        encoder_input_ids: torch.LongTensor,
        prefix_allowed_tokens_fn: Callable[[int, torch.Tensor], List[int]],
        logits_processor: Optional[LogitsProcessorList],
    ) -> LogitsProcessorList:
        """
        This class returns a [`LogitsProcessorList`] list object that contains all relevant [`LogitsProcessor`]
        instances used to modify the scores of the language model head.
        """
        # instantiate processors list
845
        processors = LogitsProcessorList()
846

847
848
849
        if generation_config.sequence_bias is not None:
            processors.append(SequenceBiasLogitsProcessor(sequence_bias=generation_config.sequence_bias))

850
        if generation_config.diversity_penalty is not None and generation_config.diversity_penalty > 0.0:
851
852
            processors.append(
                HammingDiversityLogitsProcessor(
853
854
855
                    diversity_penalty=generation_config.diversity_penalty,
                    num_beams=generation_config.num_beams,
                    num_beam_groups=generation_config.num_beam_groups,
856
857
                )
            )
Karim Foda's avatar
Karim Foda committed
858
859
860
861
862
863
864
865
866
        if (
            generation_config.encoder_repetition_penalty is not None
            and generation_config.encoder_repetition_penalty != 1.0
        ):
            processors.append(
                EncoderRepetitionPenaltyLogitsProcessor(
                    penalty=generation_config.encoder_repetition_penalty, encoder_input_ids=encoder_input_ids
                )
            )
867
868
869
870
871
872
873
874
        if generation_config.repetition_penalty is not None and generation_config.repetition_penalty != 1.0:
            processors.append(RepetitionPenaltyLogitsProcessor(penalty=generation_config.repetition_penalty))
        if generation_config.no_repeat_ngram_size is not None and generation_config.no_repeat_ngram_size > 0:
            processors.append(NoRepeatNGramLogitsProcessor(generation_config.no_repeat_ngram_size))
        if (
            generation_config.encoder_no_repeat_ngram_size is not None
            and generation_config.encoder_no_repeat_ngram_size > 0
        ):
875
            if self.config.is_encoder_decoder:
876
877
878
879
880
                processors.append(
                    EncoderNoRepeatNGramLogitsProcessor(
                        generation_config.encoder_no_repeat_ngram_size, encoder_input_ids
                    )
                )
881
882
883
884
            else:
                raise ValueError(
                    "It's impossible to use `encoder_no_repeat_ngram_size` with decoder-only architecture"
                )
885
886
887
888
889
890
891
892
893
894
        if generation_config.bad_words_ids is not None:
            processors.append(
                NoBadWordsLogitsProcessor(generation_config.bad_words_ids, generation_config.eos_token_id)
            )
        if (
            generation_config.min_length is not None
            and generation_config.eos_token_id is not None
            and generation_config.min_length > 0
        ):
            processors.append(MinLengthLogitsProcessor(generation_config.min_length, generation_config.eos_token_id))
895
896
897
898
899
900
901
902
903
904
        if (
            generation_config.min_new_tokens is not None
            and generation_config.eos_token_id is not None
            and generation_config.min_new_tokens > 0
        ):
            processors.append(
                MinNewTokensLengthLogitsProcessor(
                    input_ids_seq_length, generation_config.min_new_tokens, generation_config.eos_token_id
                )
            )
905
        if prefix_allowed_tokens_fn is not None:
906
907
908
909
910
911
912
913
914
915
916
917
            processors.append(
                PrefixConstrainedLogitsProcessor(
                    prefix_allowed_tokens_fn, generation_config.num_beams // generation_config.num_beam_groups
                )
            )
        if generation_config.forced_bos_token_id is not None:
            processors.append(ForcedBOSTokenLogitsProcessor(generation_config.forced_bos_token_id))
        if generation_config.forced_eos_token_id is not None:
            processors.append(
                ForcedEOSTokenLogitsProcessor(generation_config.max_length, generation_config.forced_eos_token_id)
            )
        if generation_config.remove_invalid_values is True:
918
            processors.append(InfNanRemoveLogitsProcessor())
919
        if generation_config.exponential_decay_length_penalty is not None:
920
            processors.append(
921
922
923
                ExponentialDecayLengthPenalty(
                    generation_config.exponential_decay_length_penalty,
                    generation_config.eos_token_id,
924
                    input_ids_seq_length,
925
                )
926
            )
927
928
929
        if generation_config.suppress_tokens is not None:
            processors.append(SuppressTokensLogitsProcessor(generation_config.suppress_tokens))
        if generation_config.begin_suppress_tokens is not None:
930
            begin_index = input_ids_seq_length
931
932
933
934
935
936
937
938
939
940
941
942
943
            begin_index = (
                begin_index
                if (input_ids_seq_length > 1 or generation_config.forced_bos_token_id is None)
                else begin_index + 1
            )
            if generation_config.forced_decoder_ids is not None:
                # generation starts after the last token that is forced
                begin_index += generation_config.forced_decoder_ids[-1][0]
            processors.append(
                SuppressTokensAtBeginLogitsProcessor(generation_config.begin_suppress_tokens, begin_index)
            )
        if generation_config.forced_decoder_ids is not None:
            processors.append(ForceTokensLogitsProcessor(generation_config.forced_decoder_ids))
Sanchit Gandhi's avatar
Sanchit Gandhi committed
944
945
        if generation_config.guidance_scale is not None and generation_config.guidance_scale > 1:
            processors.append(ClassifierFreeGuidanceLogitsProcessor(generation_config.guidance_scale))
946
947
        processors = self._merge_criteria_processor_list(processors, logits_processor)
        # `LogitNormalization` should always be the last logit processor, when present
948
        if generation_config.renormalize_logits is True:
949
950
951
952
            processors.append(LogitNormalization())
        return processors

    def _get_stopping_criteria(
953
        self, generation_config: GenerationConfig, stopping_criteria: Optional[StoppingCriteriaList]
954
955
    ) -> StoppingCriteriaList:
        criteria = StoppingCriteriaList()
956
        if generation_config.max_length is not None:
957
958
959
960
961
962
963
            max_position_embeddings = getattr(self.config, "max_position_embeddings", None)
            criteria.append(
                MaxLengthCriteria(
                    max_length=generation_config.max_length,
                    max_position_embeddings=max_position_embeddings,
                )
            )
964
965
        if generation_config.max_time is not None:
            criteria.append(MaxTimeCriteria(max_time=generation_config.max_time))
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
        criteria = self._merge_criteria_processor_list(criteria, stopping_criteria)
        return criteria

    def _merge_criteria_processor_list(
        self,
        default_list: Union[LogitsProcessorList, StoppingCriteriaList],
        custom_list: Union[LogitsProcessorList, StoppingCriteriaList],
    ) -> Union[LogitsProcessorList, StoppingCriteriaList]:
        if len(custom_list) == 0:
            return default_list
        for default in default_list:
            for custom in custom_list:
                if type(custom) is type(default):
                    object_type = "stopping criteria" if isinstance(custom, StoppingCriteria) else "logits processor"
                    raise ValueError(
                        f"A custom {object_type} of type {type(custom)} with values {custom} has been passed to"
982
                        f" `.generate()`, but it has already been created with the values {default}. {default} has been"
983
984
                        " created by passing the corresponding arguments to generate or by the model's config default"
                        f" values. If you just want to change the default values of {object_type} consider passing"
985
                        f" them as arguments to `.generate()` instead of using a custom {object_type}."
986
987
988
989
                    )
        default_list.extend(custom_list)
        return default_list

990
    def compute_transition_scores(
991
992
993
        self,
        sequences: torch.Tensor,
        scores: Tuple[torch.Tensor],
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
        beam_indices: Optional[torch.Tensor] = None,
        normalize_logits: bool = False,
    ) -> torch.Tensor:
        """
        Computes the transition scores of sequences given the generation scores (and beam indices, if beam search was
        used). This is a convenient method to quicky obtain the scores of the selected tokens at generation time.

        Parameters:
            sequences (`torch.LongTensor`):
                The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or
                shorter if all batches finished early due to the `eos_token_id`.
            scores (`tuple(torch.FloatTensor)`):
                Transition scores for each vocabulary token at each generation step. Beam transition scores consisting
                of log probabilities of tokens conditioned on log softmax of previously generated tokens Tuple of
                `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with
                each tensor of shape `(batch_size*num_beams, config.vocab_size)`.
1010
            beam_indices (`torch.LongTensor`, *optional*):
1011
                Beam indices of generated token id at each generation step. `torch.LongTensor` of shape
1012
                `(batch_size*num_return_sequences, sequence_length)`. Only required if a `num_beams>1` at
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
                generate-time.
            normalize_logits (`bool`, *optional*, defaults to `False`):
                Whether to normalize the logits (which, for legacy reasons, may be unnormalized).

        Return:
            `torch.Tensor`: A `torch.Tensor` of shape `(batch_size*num_return_sequences, sequence_length)` containing
                the transition scores (logits)

        Examples:

        ```python
        >>> from transformers import GPT2Tokenizer, AutoModelForCausalLM
        >>> import numpy as np

        >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")
        >>> tokenizer.pad_token_id = tokenizer.eos_token_id
        >>> inputs = tokenizer(["Today is"], return_tensors="pt")

        >>> # Example 1: Print the scores for each token generated with Greedy Search
        >>> outputs = model.generate(**inputs, max_new_tokens=5, return_dict_in_generate=True, output_scores=True)
        >>> transition_scores = model.compute_transition_scores(
        ...     outputs.sequences, outputs.scores, normalize_logits=True
        ... )
1037
1038
1039
        >>> # input_length is the length of the input prompt for decoder-only models, like the GPT family, and 1 for
        >>> # encoder-decoder models, like BART or T5.
        >>> input_length = 1 if model.config.is_encoder_decoder else inputs.input_ids.shape[1]
1040
1041
1042
        >>> generated_tokens = outputs.sequences[:, input_length:]
        >>> for tok, score in zip(generated_tokens[0], transition_scores[0]):
        ...     # | token | token string | logits | probability
1043
1044
1045
1046
1047
1048
        ...     print(f"| {tok:5d} | {tokenizer.decode(tok):8s} | {score.numpy():.3f} | {np.exp(score.numpy()):.2%}")
        |   262 |  the     | -1.414 | 24.33%
        |  1110 |  day     | -2.609 | 7.36%
        |   618 |  when    | -2.010 | 13.40%
        |   356 |  we      | -1.859 | 15.58%
        |   460 |  can     | -2.508 | 8.14%
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062

        >>> # Example 2: Reconstruct the sequence scores from Beam Search
        >>> outputs = model.generate(
        ...     **inputs,
        ...     max_new_tokens=5,
        ...     num_beams=4,
        ...     num_return_sequences=4,
        ...     return_dict_in_generate=True,
        ...     output_scores=True,
        ... )
        >>> transition_scores = model.compute_transition_scores(
        ...     outputs.sequences, outputs.scores, outputs.beam_indices, normalize_logits=False
        ... )
        >>> # If you sum the generated tokens' scores and apply the length penalty, you'll get the sequence scores.
1063
1064
1065
        >>> # Tip: recomputing the scores is only guaranteed to match with `normalize_logits=False`. Depending on the
        >>> # use case, you might want to recompute it with `normalize_logits=True`.
        >>> output_length = input_length + np.sum(transition_scores.numpy() < 0, axis=1)
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
        >>> length_penalty = model.generation_config.length_penalty
        >>> reconstructed_scores = transition_scores.sum(axis=1) / (output_length**length_penalty)
        >>> print(np.allclose(outputs.sequences_scores, reconstructed_scores))
        True
        ```"""
        # 1. In absence of `beam_indices`, we can assume that we come from e.g. greedy search, which is equivalent
        # to a beam search approach were the first (and only) beam is always selected
        if beam_indices is None:
            beam_indices = torch.arange(scores[0].shape[0]).view(-1, 1).to(sequences.device)
            beam_indices = beam_indices.expand(-1, len(scores))

        # 2. reshape scores as [batch_size*vocab_size, # generation steps] with # generation steps being
1078
1079
1080
        # seq_len - input_length
        scores = torch.stack(scores).reshape(len(scores), -1).transpose(0, 1)

1081
1082
1083
1084
1085
1086
1087
        # 3. Optionally normalize the logits (across the vocab dimension)
        if normalize_logits:
            scores = scores.reshape(-1, self.config.vocab_size, scores.shape[-1])
            scores = torch.nn.functional.log_softmax(scores, dim=1)
            scores = scores.reshape(-1, scores.shape[-1])

        # 4. cut beam_indices to longest beam length
1088
1089
        beam_indices_mask = beam_indices < 0
        max_beam_length = (1 - beam_indices_mask.long()).sum(-1).max()
1090
        beam_indices = beam_indices.clone()[:, :max_beam_length]
1091
1092
        beam_indices_mask = beam_indices_mask[:, :max_beam_length]

1093
        # 5. Set indices of beams that finished early to 0; such indices will be masked correctly afterwards
1094
1095
        beam_indices[beam_indices_mask] = 0

1096
        # 6. multiply beam_indices with vocab size to gather correctly from scores
1097
1098
        beam_sequence_indices = beam_indices * self.config.vocab_size

1099
        # 7. Define which indices contributed to scores
1100
1101
1102
        cut_idx = sequences.shape[-1] - max_beam_length
        indices = sequences[:, cut_idx:] + beam_sequence_indices

1103
        # 8. Compute scores
1104
1105
        transition_scores = scores.gather(0, indices)

1106
        # 9. Mask out transition_scores of beams that stopped early
1107
1108
1109
1110
1111
1112
1113
1114
1115
        transition_scores[beam_indices_mask] = 0

        return transition_scores

    def _validate_model_class(self):
        """
        Confirms that the model class is compatible with generation. If not, raises an exception that points to the
        right class to use.
        """
1116
        if not self.can_generate():
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
            generate_compatible_mappings = [
                MODEL_FOR_CAUSAL_LM_MAPPING,
                MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING,
                MODEL_FOR_VISION_2_SEQ_MAPPING,
                MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
                MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
            ]
            generate_compatible_classes = set()
            for model_mapping in generate_compatible_mappings:
                supported_models = model_mapping.get(type(self.config), default=None)
                if supported_models is not None:
                    generate_compatible_classes.add(supported_models.__name__)
            exception_message = (
                f"The current model class ({self.__class__.__name__}) is not compatible with `.generate()`, as "
                "it doesn't have a language model head."
            )
            if generate_compatible_classes:
                exception_message += f" Please use one of the following classes instead: {generate_compatible_classes}"
            raise TypeError(exception_message)

    def _validate_model_kwargs(self, model_kwargs: Dict[str, Any]):
        """Validates model kwargs for generation. Generate argument typos will also be caught here."""
        # Excludes arguments that are handled before calling any model function
        if self.config.is_encoder_decoder:
            for key in ["decoder_input_ids"]:
                model_kwargs.pop(key, None)

        unused_model_args = []
        model_args = set(inspect.signature(self.prepare_inputs_for_generation).parameters)
1146
1147
1148
        # `kwargs`/`model_kwargs` is often used to handle optional forward pass inputs like `attention_mask`. If
        # `prepare_inputs_for_generation` doesn't accept them, then a stricter check can be made ;)
        if "kwargs" in model_args or "model_kwargs" in model_args:
1149
            model_args |= set(inspect.signature(self.forward).parameters)
1150
1151
1152
1153
1154
1155
1156

        # Encoder-Decoder models may also need Encoder arguments from `model_kwargs`
        if self.config.is_encoder_decoder:
            base_model = getattr(self, self.base_model_prefix, None)

            # allow encoder kwargs
            encoder = getattr(self, "encoder", None)
1157
1158
1159
1160
1161
            # `MusicgenForConditionalGeneration` has `text_encoder` and `audio_encoder`.
            # Also, it has `base_model_prefix = "encoder_decoder"` but there is no `self.encoder_decoder`
            # TODO: A better way to handle this.
            if encoder is None and base_model is not None:
                encoder = getattr(base_model, "encoder", None)
1162

1163
1164
1165
            if encoder is not None:
                encoder_model_args = set(inspect.signature(encoder.forward).parameters)
                model_args |= encoder_model_args
1166
1167
1168

            # allow decoder kwargs
            decoder = getattr(self, "decoder", None)
1169
1170
            if decoder is None and base_model is not None:
                decoder = getattr(base_model, "decoder", None)
1171

1172
1173
1174
            if decoder is not None:
                decoder_model_args = set(inspect.signature(decoder.forward).parameters)
                model_args |= {f"decoder_{x}" for x in decoder_model_args}
1175

1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
        for key, value in model_kwargs.items():
            if value is not None and key not in model_args:
                unused_model_args.append(key)

        if unused_model_args:
            raise ValueError(
                f"The following `model_kwargs` are not used by the model: {unused_model_args} (note: typos in the"
                " generate arguments will also show up in this list)"
            )

    @torch.no_grad()
    def generate(
        self,
        inputs: Optional[torch.Tensor] = None,
1190
        generation_config: Optional[GenerationConfig] = None,
1191
1192
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
1193
        prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
1194
        synced_gpus: Optional[bool] = None,
1195
        assistant_model: Optional["PreTrainedModel"] = None,
1196
        streamer: Optional["BaseStreamer"] = None,
1197
        **kwargs,
1198
1199
1200
    ) -> Union[GenerateOutput, torch.LongTensor]:
        r"""

1201
        Generates sequences of token ids for models with a language modeling head.
1202
1203
1204

        <Tip warning={true}>

1205
1206
        Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
        model's default generation configuration. You can override any `generation_config` by passing the corresponding
1207
        parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
1208

1209
        For an overview of generation strategies and code examples, check out the [following
1210
        guide](../generation_strategies).
1211

1212
        </Tip>
1213
1214
1215
1216
1217
1218
1219

        Parameters:
            inputs (`torch.Tensor` of varying shape depending on the modality, *optional*):
                The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the
                method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs`
                should of in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of
                `input_ids`, `input_values`, `input_features`, or `pixel_values`.
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
            generation_config (`~generation.GenerationConfig`, *optional*):
                The generation configuration to be used as base parametrization for the generation call. `**kwargs`
                passed to generate matching the attributes of `generation_config` will override them. If
                `generation_config` is not provided, the default will be used, which had the following loading
                priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
                configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
                default values, whose documentation should be checked to parameterize generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                Custom logits processors that complement the default logits processors built from arguments and
                generation config. If a logit processor is passed that is already created with the arguments or a
                generation config an error is thrown. This feature is intended for advanced users.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                Custom stopping criteria that complement the default stopping criteria built from arguments and a
                generation config. If a stopping criteria is passed that is already created with the arguments or a
                generation config an error is thrown. This feature is intended for advanced users.
1235
1236
1237
1238
1239
1240
1241
            prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
                If provided, this function constraints the beam search to allowed tokens only at each step. If not
                provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
                `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
                on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
                for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
                Retrieval](https://arxiv.org/abs/2010.00904).
1242
1243
1244
1245
            synced_gpus (`bool`, *optional*):
                Whether to continue running the while loop until max_length. Unless overridden this flag will be set to
                `True` under DeepSpeed ZeRO Stage 3 multiple GPUs environment to avoid hanging if one GPU finished
                generating before other GPUs. Otherwise it'll be set to `False`.
1246
1247
1248
1249
1250
            assistant_model (`PreTrainedModel`, *optional*):
                An assistant model that can be used to accelerate generation. The assistant model must have the exact
                same tokenizer. The acceleration is achieved when forecasting candidate tokens with the assistent model
                is much faster than running generation with the model you're calling generate from. As such, the
                assistant model should be much smaller.
1251
1252
1253
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
1254
            kwargs (`Dict[str, Any]`, *optional*):
1255
1256
1257
                Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
                forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
                specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277

        Return:
            [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
            or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor`.

                If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible
                [`~utils.ModelOutput`] types are:

                    - [`~generation.GreedySearchDecoderOnlyOutput`],
                    - [`~generation.SampleDecoderOnlyOutput`],
                    - [`~generation.BeamSearchDecoderOnlyOutput`],
                    - [`~generation.BeamSampleDecoderOnlyOutput`]

                If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible
                [`~utils.ModelOutput`] types are:

                    - [`~generation.GreedySearchEncoderDecoderOutput`],
                    - [`~generation.SampleEncoderDecoderOutput`],
                    - [`~generation.BeamSearchEncoderDecoderOutput`],
                    - [`~generation.BeamSampleEncoderDecoderOutput`]
1278
        """
1279
1280

        if synced_gpus is None:
1281
            if is_deepspeed_zero3_enabled() and dist.get_world_size() > 1:
1282
1283
1284
1285
                synced_gpus = True
            else:
                synced_gpus = False

1286
        # 1. Handle `generation_config` and kwargs that might update it, and validate the `.generate()` call
1287
        self._validate_model_class()
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299

        # priority: `generation_config` argument > `model.generation_config` (the default generation config)
        if generation_config is None:
            # legacy: users may modify the model configuration to control generation -- update the generation config
            # model attribute accordingly, if it was created from the model config
            if self.generation_config._from_model_config:
                new_generation_config = GenerationConfig.from_model_config(self.config)
                if new_generation_config != self.generation_config:
                    warnings.warn(
                        "You have modified the pretrained model configuration to control generation. This is a"
                        " deprecated strategy to control generation and will be removed soon, in a future version."
                        " Please use a generation configuration file (see"
1300
                        " https://huggingface.co/docs/transformers/main_classes/text_generation )"
1301
1302
1303
1304
1305
1306
                    )
                    self.generation_config = new_generation_config
            generation_config = self.generation_config

        generation_config = copy.deepcopy(generation_config)
        model_kwargs = generation_config.update(**kwargs)  # All unused kwargs must be model kwargs
1307
        generation_config.validate()
1308
1309
        self._validate_model_kwargs(model_kwargs.copy())

1310
        # 2. Set generation parameters if not already defined
1311
1312
1313
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()

1314
        if generation_config.pad_token_id is None and generation_config.eos_token_id is not None:
1315
1316
1317
1318
1319
            if model_kwargs.get("attention_mask", None) is None:
                logger.warning(
                    "The attention mask and the pad token id were not set. As a consequence, you may observe "
                    "unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results."
                )
1320
1321
1322
1323
1324
            eos_token_id = generation_config.eos_token_id
            if isinstance(eos_token_id, list):
                eos_token_id = eos_token_id[0]
            logger.warning(f"Setting `pad_token_id` to `eos_token_id`:{eos_token_id} for open-end generation.")
            generation_config.pad_token_id = eos_token_id
1325

1326
        # 3. Define model inputs
1327
1328
1329
1330
        # inputs_tensor has to be defined
        # model_input_name is defined if model-specific keyword input is passed
        # otherwise model_input_name is None
        # all model-specific keyword inputs are removed from `model_kwargs`
1331
1332
1333
        inputs_tensor, model_input_name, model_kwargs = self._prepare_model_inputs(
            inputs, generation_config.bos_token_id, model_kwargs
        )
1334
1335
        batch_size = inputs_tensor.shape[0]

1336
1337
1338
        # 4. Define other model kwargs
        model_kwargs["output_attentions"] = generation_config.output_attentions
        model_kwargs["output_hidden_states"] = generation_config.output_hidden_states
1339
1340
1341
1342
1343
1344
        # decoder-only models with inputs_embeds forwarding must use caching (otherwise we can't detect whether we are
        # generating the first new token or not, and we only want to use the embeddings for the first new token)
        if not self.config.is_encoder_decoder and model_input_name == "inputs_embeds":
            model_kwargs["use_cache"] = True
        else:
            model_kwargs["use_cache"] = generation_config.use_cache
1345
1346
1347
1348
1349
1350

        accepts_attention_mask = "attention_mask" in set(inspect.signature(self.forward).parameters.keys())
        requires_attention_mask = "encoder_outputs" not in model_kwargs

        if model_kwargs.get("attention_mask", None) is None and requires_attention_mask and accepts_attention_mask:
            model_kwargs["attention_mask"] = self._prepare_attention_mask_for_generation(
1351
                inputs_tensor, generation_config.pad_token_id, generation_config.eos_token_id
1352
1353
1354
1355
            )

        # decoder-only models should use left-padding for generation
        if not self.config.is_encoder_decoder:
1356
1357
            # If `input_ids` was given, check if the last id in any sequence is `pad_token_id`
            # Note: If using, `inputs_embeds` this check does not work, because we want to be more hands-off.
1358
1359
            if (
                generation_config.pad_token_id is not None
1360
                and len(inputs_tensor.shape) == 2
1361
1362
                and torch.sum(inputs_tensor[:, -1] == generation_config.pad_token_id) > 0
            ):
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
                logger.warning(
                    "A decoder-only architecture is being used, but right-padding was detected! For correct "
                    "generation results, please set `padding_side='left'` when initializing the tokenizer."
                )

        if self.config.is_encoder_decoder and "encoder_outputs" not in model_kwargs:
            # if model is encoder decoder encoder_outputs are created
            # and added to `model_kwargs`
            model_kwargs = self._prepare_encoder_decoder_kwargs_for_generation(
                inputs_tensor, model_kwargs, model_input_name
            )

1375
        # 5. Prepare `input_ids` which will be used for auto-regressive generation
1376
        if self.config.is_encoder_decoder:
1377
1378
1379
1380
            input_ids, model_kwargs = self._prepare_decoder_input_ids_for_generation(
                batch_size=batch_size,
                model_input_name=model_input_name,
                model_kwargs=model_kwargs,
1381
1382
                decoder_start_token_id=generation_config.decoder_start_token_id,
                bos_token_id=generation_config.bos_token_id,
1383
1384
1385
                device=inputs_tensor.device,
            )
        else:
1386
            input_ids = inputs_tensor if model_input_name == "input_ids" else model_kwargs.pop("input_ids")
1387

1388
1389
1390
        if streamer is not None:
            streamer.put(input_ids.cpu())

1391
        # 6. Prepare `max_length` depending on other stopping criteria.
1392
        input_ids_seq_length = input_ids.shape[-1]
1393
        has_default_max_length = kwargs.get("max_length") is None and generation_config.max_length is not None
1394
1395
        if has_default_max_length and generation_config.max_new_tokens is None and generation_config.max_length != 20:
            # 20 is the default max_length of the generation config
1396
            warnings.warn(
1397
1398
                f"Using the model-agnostic default `max_length` (={generation_config.max_length}) "
                "to control the generation length.  recommend setting `max_new_tokens` to control the maximum length of the generation.",
1399
1400
                UserWarning,
            )
1401
1402
        elif generation_config.max_new_tokens is not None:
            if not has_default_max_length:
1403
                logger.warning(
1404
1405
1406
                    f"Both `max_new_tokens` (={generation_config.max_new_tokens}) and `max_length`(="
                    f"{generation_config.max_length}) seem to have been set. `max_new_tokens` will take precedence. "
                    "Please refer to the documentation for more information. "
1407
                    "(https://huggingface.co/docs/transformers/main/en/main_classes/text_generation)"
1408
                )
1409
            generation_config.max_length = generation_config.max_new_tokens + input_ids_seq_length
1410

1411
        if generation_config.min_length is not None and generation_config.min_length > generation_config.max_length:
1412
            raise ValueError(
1413
1414
                f"Unfeasible length constraints: the minimum length ({generation_config.min_length}) is larger than"
                f" the maximum length ({generation_config.max_length})"
1415
            )
1416
        if input_ids_seq_length >= generation_config.max_length:
1417
1418
1419
            input_ids_string = "decoder_input_ids" if self.config.is_encoder_decoder else "input_ids"
            logger.warning(
                f"Input length of {input_ids_string} is {input_ids_seq_length}, but `max_length` is set to"
1420
1421
                f" {generation_config.max_length}. This can lead to unexpected behavior. You should consider"
                " increasing `max_new_tokens`."
1422
1423
            )

1424
1425
1426
1427
        # 7. determine generation mode
        is_constraint_gen_mode = (
            generation_config.constraints is not None or generation_config.force_words_ids is not None
        )
1428
1429

        is_contrastive_search_gen_mode = (
1430
1431
            (generation_config.num_beams == 1)
            and generation_config.top_k is not None
1432
1433
1434
1435
            and generation_config.top_k > 1
            and generation_config.do_sample is False
            and generation_config.penalty_alpha is not None
            and generation_config.penalty_alpha > 0
1436
1437
1438
        )

        is_greedy_gen_mode = (
1439
1440
1441
            (generation_config.num_beams == 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is False
1442
1443
1444
1445
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_sample_gen_mode = (
1446
1447
1448
            (generation_config.num_beams == 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is True
1449
1450
1451
1452
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_beam_gen_mode = (
1453
1454
1455
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is False
1456
1457
1458
1459
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_beam_sample_gen_mode = (
1460
1461
1462
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups == 1)
            and generation_config.do_sample is True
1463
1464
1465
1466
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
        is_group_beam_gen_mode = (
1467
1468
            (generation_config.num_beams > 1)
            and (generation_config.num_beam_groups > 1)
1469
1470
1471
            and not is_constraint_gen_mode
            and not is_contrastive_search_gen_mode
        )
1472
        is_assisted_gen_mode = False
1473
        if assistant_model is not None:
1474
            if not (is_greedy_gen_mode or is_sample_gen_mode):
1475
                raise ValueError(
1476
1477
                    "You've set `assistant_model`, which triggers assisted generate. Currently, assisted generate "
                    "is only supported with Greedy Search and Sample."
1478
                )
1479
            is_assisted_gen_mode = True
1480

1481
        if generation_config.num_beam_groups > generation_config.num_beams:
1482
            raise ValueError("`num_beam_groups` has to be smaller or equal to `num_beams`")
1483
        if is_group_beam_gen_mode and generation_config.do_sample is True:
1484
1485
1486
1487
            raise ValueError(
                "Diverse beam search cannot be used in sampling mode. Make sure that `do_sample` is set to `False`."
            )

1488
1489
1490
1491
1492
        if streamer is not None and (generation_config.num_beams > 1):
            raise ValueError(
                "`streamer` cannot be used with beam search (yet!). Make sure that `num_beams` is set to 1."
            )

1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
        if self.device.type != input_ids.device.type:
            warnings.warn(
                "You are calling .generate() with the `input_ids` being on a device type different"
                f" than your model's device. `input_ids` is on {input_ids.device.type}, whereas the model"
                f" is on {self.device.type}. You may experience unexpected behaviors or slower generation."
                " Please make sure that you have put `input_ids` to the"
                f" correct device by calling for example input_ids = input_ids.to('{self.device.type}') before"
                " running `.generate()`.",
                UserWarning,
            )

1504
        # 8. prepare distribution pre_processing samplers
1505
        logits_processor = self._get_logits_processor(
1506
            generation_config=generation_config,
1507
1508
1509
1510
1511
1512
            input_ids_seq_length=input_ids_seq_length,
            encoder_input_ids=inputs_tensor,
            prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
            logits_processor=logits_processor,
        )

1513
        # 9. prepare stopping criteria
1514
        stopping_criteria = self._get_stopping_criteria(
1515
            generation_config=generation_config, stopping_criteria=stopping_criteria
1516
        )
1517
        # 10. go into different generation modes
1518
        if is_assisted_gen_mode:
1519
1520
            if generation_config.num_return_sequences > 1:
                raise ValueError(
1521
                    "num_return_sequences has to be 1 when doing assisted generate, "
1522
1523
1524
                    f"but is {generation_config.num_return_sequences}."
                )
            if batch_size > 1:
1525
                raise ValueError("assisted generate is only supported for batch_size = 1")
1526
            if not model_kwargs["use_cache"]:
1527
                raise ValueError("assisted generate requires `use_cache=True`")
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539

            # 11. If the assistant model is an encoder-decoder, prepare its encoder outputs
            if assistant_model.config.is_encoder_decoder:
                assistant_model_kwargs = copy.deepcopy(model_kwargs)
                inputs_tensor, model_input_name, assistant_model_kwargs = assistant_model._prepare_model_inputs(
                    inputs_tensor, assistant_model.generation_config.bos_token_id, assistant_model_kwargs
                )
                assistant_model_kwargs = assistant_model._prepare_encoder_decoder_kwargs_for_generation(
                    inputs_tensor, assistant_model_kwargs, model_input_name
                )
                model_kwargs["assistant_encoder_outputs"] = assistant_model_kwargs["encoder_outputs"]

1540
1541
            # 12. run assisted generate
            return self.assisted_decoding(
1542
1543
                input_ids,
                assistant_model=assistant_model,
1544
                do_sample=generation_config.do_sample,
1545
                logits_processor=logits_processor,
1546
                logits_warper=self._get_logits_warper(generation_config) if generation_config.do_sample else None,
1547
1548
1549
1550
1551
1552
1553
1554
1555
                stopping_criteria=stopping_criteria,
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
                synced_gpus=synced_gpus,
                streamer=streamer,
                **model_kwargs,
            )
1556
        if is_greedy_gen_mode:
1557
            if generation_config.num_return_sequences > 1:
1558
                raise ValueError(
1559
1560
                    "num_return_sequences has to be 1 when doing greedy search, "
                    f"but is {generation_config.num_return_sequences}."
1561
1562
                )

1563
            # 11. run greedy search
1564
1565
1566
1567
            return self.greedy_search(
                input_ids,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
1568
1569
1570
1571
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1572
                synced_gpus=synced_gpus,
1573
                streamer=streamer,
1574
1575
1576
1577
                **model_kwargs,
            )

        elif is_contrastive_search_gen_mode:
1578
            if generation_config.num_return_sequences > 1:
1579
                raise ValueError(
1580
1581
                    "num_return_sequences has to be 1 when doing contrastive search, "
                    f"but is {generation_config.num_return_sequences}."
1582
                )
1583
1584
            if not model_kwargs["use_cache"]:
                raise ValueError("Contrastive search requires `use_cache=True`")
1585
1586
1587

            return self.contrastive_search(
                input_ids,
1588
1589
                top_k=generation_config.top_k,
                penalty_alpha=generation_config.penalty_alpha,
1590
1591
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
1592
1593
1594
1595
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1596
                synced_gpus=synced_gpus,
1597
                streamer=streamer,
1598
                sequential=generation_config.low_memory,
1599
1600
1601
1602
                **model_kwargs,
            )

        elif is_sample_gen_mode:
1603
1604
            # 11. prepare logits warper
            logits_warper = self._get_logits_warper(generation_config)
1605

1606
            # 12. expand input_ids with `num_return_sequences` additional sequences per batch
1607
1608
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
1609
                expand_size=generation_config.num_return_sequences,
1610
1611
1612
1613
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )

1614
            # 13. run sample
1615
1616
1617
1618
1619
            return self.sample(
                input_ids,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                stopping_criteria=stopping_criteria,
1620
1621
1622
1623
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1624
                synced_gpus=synced_gpus,
1625
                streamer=streamer,
1626
1627
1628
1629
                **model_kwargs,
            )

        elif is_beam_gen_mode:
1630
            if generation_config.num_return_sequences > generation_config.num_beams:
1631
1632
1633
1634
1635
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

1636
            # 11. prepare beam search scorer
1637
1638
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size,
1639
                num_beams=generation_config.num_beams,
1640
                device=inputs_tensor.device,
1641
1642
1643
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
1644
                max_length=generation_config.max_length,
1645
            )
1646
            # 12. interleave input_ids with `num_beams` additional sequences per batch
1647
1648
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
1649
                expand_size=generation_config.num_beams,
1650
1651
1652
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
1653
            # 13. run beam search
1654
1655
1656
1657
1658
            return self.beam_search(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
1659
1660
1661
1662
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1663
1664
1665
1666
1667
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_beam_sample_gen_mode:
1668
1669
            # 11. prepare logits warper
            logits_warper = self._get_logits_warper(generation_config)
1670
1671
1672

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")
1673
            # 12. prepare beam search scorer
1674
            beam_scorer = BeamSearchScorer(
1675
1676
                batch_size=batch_size * generation_config.num_return_sequences,
                num_beams=generation_config.num_beams,
1677
                device=inputs_tensor.device,
1678
1679
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
1680
                max_length=generation_config.max_length,
1681
1682
            )

1683
            # 13. interleave input_ids with `num_beams` additional sequences per batch
1684
1685
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
1686
                expand_size=generation_config.num_beams * generation_config.num_return_sequences,
1687
1688
1689
1690
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )

1691
            # 14. run beam sample
1692
1693
1694
1695
1696
1697
            return self.beam_sample(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                stopping_criteria=stopping_criteria,
1698
1699
1700
1701
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1702
1703
1704
1705
1706
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_group_beam_gen_mode:
1707
            if generation_config.num_return_sequences > generation_config.num_beams:
1708
1709
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

1710
            if generation_config.num_beams % generation_config.num_beam_groups != 0:
1711
1712
                raise ValueError("`num_beams` should be divisible by `num_beam_groups` for group beam search.")

1713
1714
1715
1716
1717
            if generation_config.diversity_penalty == 0.0:
                raise ValueError(
                    "`diversity_penalty` should be greater than `0.0`, otherwise your beam groups will be identical."
                )

1718
1719
1720
            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

1721
1722
            has_default_typical_p = kwargs.get("typical_p") is None and generation_config.typical_p == 1.0
            if not has_default_typical_p:
1723
1724
                raise ValueError("Decoder argument `typical_p` is not supported with beam groups.")

1725
            # 11. prepare beam search scorer
1726
1727
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size,
1728
                num_beams=generation_config.num_beams,
1729
                device=inputs_tensor.device,
1730
1731
1732
1733
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
                num_beam_groups=generation_config.num_beam_groups,
1734
                max_length=generation_config.max_length,
1735
            )
1736
            # 12. interleave input_ids with `num_beams` additional sequences per batch
1737
1738
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
1739
                expand_size=generation_config.num_beams,
1740
1741
1742
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
1743
            # 13. run beam search
1744
1745
1746
1747
1748
            return self.group_beam_search(
                input_ids,
                beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
1749
1750
1751
1752
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1753
1754
1755
1756
1757
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

        elif is_constraint_gen_mode:
1758
            if generation_config.num_return_sequences > generation_config.num_beams:
1759
1760
1761
1762
1763
                raise ValueError("`num_return_sequences` has to be smaller or equal to `num_beams`.")

            if stopping_criteria.max_length is None:
                raise ValueError("`max_length` needs to be a stopping_criteria for now.")

1764
            if generation_config.num_beams <= 1:
1765
1766
                raise ValueError("`num_beams` needs to be greater than 1 for constrained generation.")

1767
            if generation_config.do_sample:
1768
1769
                raise ValueError("`do_sample` needs to be false for constrained generation.")

1770
            if generation_config.num_beam_groups is not None and generation_config.num_beam_groups > 1:
1771
1772
1773
                raise ValueError("`num_beam_groups` not supported yet for constrained generation.")

            final_constraints = []
1774
1775
            if generation_config.constraints is not None:
                final_constraints = generation_config.constraints
1776

1777
            if generation_config.force_words_ids is not None:
1778
1779
1780
1781

                def typeerror():
                    raise ValueError(
                        "`force_words_ids` has to either be a `List[List[List[int]]]` or `List[List[int]]`"
1782
                        f"of positive integers, but is {generation_config.force_words_ids}."
1783
1784
                    )

1785
1786
1787
1788
                if (
                    not isinstance(generation_config.force_words_ids, list)
                    or len(generation_config.force_words_ids) == 0
                ):
1789
1790
                    typeerror()

1791
                for word_ids in generation_config.force_words_ids:
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
                    if isinstance(word_ids[0], list):
                        if not isinstance(word_ids, list) or len(word_ids) == 0:
                            typeerror()
                        if any(not isinstance(token_ids, list) for token_ids in word_ids):
                            typeerror()
                        if any(
                            any((not isinstance(token_id, int) or token_id < 0) for token_id in token_ids)
                            for token_ids in word_ids
                        ):
                            typeerror()

                        constraint = DisjunctiveConstraint(word_ids)
                    else:
                        if not isinstance(word_ids, list) or len(word_ids) == 0:
                            typeerror()
                        if any((not isinstance(token_id, int) or token_id < 0) for token_id in word_ids):
                            typeerror()

                        constraint = PhrasalConstraint(word_ids)
                    final_constraints.append(constraint)

1813
            # 11. prepare beam search scorer
1814
1815
1816
            constrained_beam_scorer = ConstrainedBeamSearchScorer(
                constraints=final_constraints,
                batch_size=batch_size,
1817
                num_beams=generation_config.num_beams,
1818
                device=inputs_tensor.device,
1819
1820
1821
                length_penalty=generation_config.length_penalty,
                do_early_stopping=generation_config.early_stopping,
                num_beam_hyps_to_keep=generation_config.num_return_sequences,
1822
                max_length=generation_config.max_length,
1823
            )
1824
            # 12. interleave input_ids with `num_beams` additional sequences per batch
1825
1826
            input_ids, model_kwargs = self._expand_inputs_for_generation(
                input_ids=input_ids,
1827
                expand_size=generation_config.num_beams,
1828
1829
1830
                is_encoder_decoder=self.config.is_encoder_decoder,
                **model_kwargs,
            )
1831
            # 13. run beam search
1832
1833
1834
1835
1836
            return self.constrained_beam_search(
                input_ids,
                constrained_beam_scorer=constrained_beam_scorer,
                logits_processor=logits_processor,
                stopping_criteria=stopping_criteria,
1837
1838
1839
1840
                pad_token_id=generation_config.pad_token_id,
                eos_token_id=generation_config.eos_token_id,
                output_scores=generation_config.output_scores,
                return_dict_in_generate=generation_config.return_dict_in_generate,
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
                synced_gpus=synced_gpus,
                **model_kwargs,
            )

    @torch.no_grad()
    def contrastive_search(
        self,
        input_ids: torch.LongTensor,
        top_k: Optional[int] = 1,
        penalty_alpha: Optional[float] = 0,
        logits_processor: Optional[LogitsProcessorList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        pad_token_id: Optional[int] = None,
1855
        eos_token_id: Optional[Union[int, List[int]]] = None,
1856
1857
1858
1859
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
1860
        synced_gpus: bool = False,
1861
        streamer: Optional["BaseStreamer"] = None,
1862
        sequential: Optional[bool] = None,
1863
1864
1865
1866
1867
1868
        **model_kwargs,
    ) -> Union[ContrastiveSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **contrastive search** and can
        be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

1869
1870
1871
1872
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.contrastive_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
1873
        guide](../generation_strategies).
1874
1875
1876

        </Tip>

1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
1894
1895
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            top_k (`int`, *optional*, defaults to 1):
                The size of the candidate set that is used to re-rank for contrastive search
            penalty_alpha (`float`, *optional*, defaults to 0):
                The degeneration penalty for contrastive search; activate when it is larger than 0
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
1896
1897
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
1910
1911
1912
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
1913
1914
            sequential (`bool`, *optional*):
                Switches topk hidden state computation from parallel to sequential to reduce memory if True.
1915
1916
1917
1918
1919
1920
1921
1922
1923
1924
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.ContrastiveSearchDecoderOnlyOutput`], [`~generation.ContrastiveSearchEncoderDecoderOutput`]
            or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.ContrastiveSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.ContrastiveSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:
        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-125m")
        >>> model = AutoModelForCausalLM.from_pretrained("facebook/opt-125m")
        >>> # set pad_token_id to eos_token_id because OPT does not have a PAD token
        >>> model.config.pad_token_id = model.config.eos_token_id
        >>> input_prompt = "DeepMind Company is"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt")
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=64)])
        >>> outputs = model.contrastive_search(
        ...     **input_ids, penalty_alpha=0.6, top_k=4, stopping_criteria=stopping_criteria
        ... )
        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['DeepMind Company is a company that focuses on the development and commercialization of artificial intelligence (AI). DeepMind’s mission is to help people understand and solve problems that are difficult to solve in the world today.\n\nIn this post, we talk about the benefits of deep learning in business and how it']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
1952
1953
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
1954
        sequential = sequential if sequential is not None else self.generation_config.low_memory
1955
1956
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
1957
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
1958
1959
1960
1961
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
1962
        output_hidden_states = (
1963
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
1964
1965
        )
        return_dict_in_generate = (
1966
1967
1968
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
1985
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002

        this_peer_finished = False  # used by synced_gpus only
        batch_size = input_ids.shape[0]

        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # if the first step in the loop, encode all the prefix and obtain: (1) past_key_values;
            # (2) last_hidden_states; (3) logit_for_next_step; (4) update model kwargs for the next step
2003
            if model_kwargs.get("past_key_values") is None:
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
                # prepare inputs
                model_kwargs["use_cache"] = True
                model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

                # encode the given prefix and prepare model inputs; encoder-decoder model process the prefix and save
                # the `encoder_outputs`
                outputs = self(
                    **model_inputs, return_dict=True, output_hidden_states=True, output_attentions=output_attentions
                )

                # last decoder hidden states will be used to compute the degeneration penalty (cosine similarity with
                # previous tokens)
                if self.config.is_encoder_decoder:
                    last_hidden_states = outputs.decoder_hidden_states[-1]
                else:
                    last_hidden_states = outputs.hidden_states[-1]
2020

2021
2022
2023
2024
                # next logit for contrastive search to select top-k candidate tokens
                logit_for_next_step = outputs.logits[:, -1, :]

                model_kwargs = self._update_model_kwargs_for_generation(
2025
2026
2027
2028
                    outputs,
                    model_kwargs,
                    is_encoder_decoder=self.config.is_encoder_decoder,
                    standardize_cache_format=True,
2029
                )
2030
2031
2032
2033
2034
                if not sequential:
                    # Expands model inputs top_k times, for batched forward passes (akin to beam search).
                    _, model_kwargs = self._expand_inputs_for_generation(
                        expand_size=top_k, is_encoder_decoder=self.config.is_encoder_decoder, **model_kwargs
                    )
2035

2036
2037
                past_key_values = model_kwargs.get("past_key_values")
                if past_key_values is None:
2038
2039
2040
2041
                    raise ValueError(
                        f"{self.__class__.__name__} does not support caching and therefore **can't** be used "
                        "for contrastive search."
                    )
2042
2043
2044
2045
                elif (
                    not isinstance(past_key_values[0], (tuple, torch.Tensor))
                    or past_key_values[0][0].shape[0] != batch_size
                ):
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
                    raise ValueError(
                        f"{self.__class__.__name__} does not have a standard cache format and therefore **can't** be "
                        "used for contrastive search without further modifications."
                    )

            # contrastive_search main logic start:
            # contrastive search decoding consists of two steps: (1) candidate tokens recall; (2) candidate re-rank by
            # degeneration penalty
            logit_for_next_step = logits_processor(input_ids, logit_for_next_step)
            logit_for_next_step = logits_warper(input_ids, logit_for_next_step)
            next_probs = nn.functional.softmax(logit_for_next_step, dim=-1)
            top_k_probs, top_k_ids = torch.topk(next_probs, dim=-1, k=top_k)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (logit_for_next_step,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # Replicates the new past_key_values to match the `top_k` candidates
            new_key_values = []
2079
            for layer in model_kwargs["past_key_values"]:
2080
2081
2082
                items = []
                # item is either the key or the value matrix
                for item in layer:
2083
2084
2085
2086
                    if sequential:
                        items.append(item.repeat_interleave(1, dim=0))
                    else:
                        items.append(item.repeat_interleave(top_k, dim=0))
2087
                new_key_values.append(items)
2088
            model_kwargs["past_key_values"] = new_key_values
2089

2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
            if sequential:
                all_outputs = {key: [] for key in outputs}  # defined in first loop iteration
                all_last_hstates, all_hstates, all_logits = [], [], []
                for i in range(top_k):
                    # compute the candidate tokens by the language model and collect their hidden_states
                    next_model_inputs = self.prepare_inputs_for_generation(top_k_ids[:, i].view(-1, 1), **model_kwargs)

                    outputs = self(
                        **next_model_inputs,
                        return_dict=True,
                        output_hidden_states=True,
                        output_attentions=output_attentions,
                    )
                    for key in all_outputs:
                        all_outputs[key].append(outputs[key])

                    if self.config.is_encoder_decoder:
                        next_hidden = outputs.decoder_hidden_states[-1]
                        full_hidden_states = outputs.decoder_hidden_states

                    else:
                        next_hidden = outputs.hidden_states[-1]
                        full_hidden_states = outputs.hidden_states

                    all_last_hstates.append(torch.squeeze(next_hidden, 0))
                    all_hstates.append(full_hidden_states)
                    all_logits.append(outputs.logits[:, -1, :])

                # stack hidden states
                next_hidden = torch.stack([all_last_hstates[i] for i in range(top_k)], dim=0)
                final_full_hstates = [0 for i in range(len(full_hidden_states))]
                for layer in range(len(full_hidden_states)):
                    final_full_hstates[layer] = torch.stack(
                        [torch.squeeze(all_hstates[i][layer], 0) for i in range(top_k)], dim=0
                    )
                full_hidden_states = tuple(final_full_hstates)

                # stack logits
                logits = torch.cat(all_logits, dim=0)
2129
2130

            else:
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
                # compute the candidate tokens by the language model and collect their hidden_states
                # assembles top_k_ids into batch of size k
                next_model_inputs = self.prepare_inputs_for_generation(top_k_ids.view(-1, 1), **model_kwargs)

                outputs = self(
                    **next_model_inputs,
                    return_dict=True,
                    output_hidden_states=True,
                    output_attentions=output_attentions,
                )
                # name is different for encoder-decoder and decoder-only models
                if self.config.is_encoder_decoder:
                    next_hidden = outputs.decoder_hidden_states[-1]
                    full_hidden_states = outputs.decoder_hidden_states
                else:
                    next_hidden = outputs.hidden_states[-1]
                    full_hidden_states = outputs.hidden_states

                logits = outputs.logits[:, -1, :]

2151
2152
2153
            context_hidden = last_hidden_states.repeat_interleave(top_k, dim=0)

            # compute the degeneration penalty and re-rank the candidates based on the degeneration penalty and the
2154
2155
            # model confidence. Keeping `selected_idx` on CPU enables multi-device contrastive search and doesn't
            # introduce (noticeable) slowdowns on single-device runs.
2156
            selected_idx = _ranking_fast(context_hidden, next_hidden, top_k_probs, penalty_alpha, top_k)
2157
            selected_idx = selected_idx.to("cpu")
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171

            # prepare for the next step: (1) next token_id; (2) past_key_values; (3) last_hidden_states for computing
            # the degeneration penalty; (4) logits for selecting next top-k candidates; (5) selected tokens scores
            # (model confidence minus degeneration penalty); (6) decoder hidden_states
            next_tokens = top_k_ids[range(len(top_k_ids)), selected_idx]
            next_hidden = torch.stack(torch.split(next_hidden.squeeze(dim=1), top_k))
            next_hidden = next_hidden[range(batch_size), selected_idx, :]
            last_hidden_states = torch.cat([last_hidden_states, next_hidden.unsqueeze(1)], dim=1)

            next_decoder_hidden_states = ()
            for layer in full_hidden_states:
                layer = torch.stack(torch.split(layer, top_k))[range(batch_size), selected_idx, :]
                next_decoder_hidden_states += (layer,)

2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
            # generate past_key_values cache of only the selected token
            if sequential:
                next_model_input = self.prepare_inputs_for_generation(
                    top_k_ids[:, selected_idx].view(-1, 1), **model_kwargs
                )

                selected_outputs = self(
                    **next_model_input,
                    return_dict=True,
                    output_hidden_states=False,
                    output_attentions=False,
                )
                next_past_key_values = selected_outputs["past_key_values"]

            else:
                next_past_key_values = self._extract_past_from_model_output(outputs, standardize_cache_format=True)
                new_key_values = ()
                for layer in next_past_key_values:
                    items = ()
                    # item is either the key or the value matrix
                    for item in layer:
                        item = torch.stack(torch.split(item, top_k, dim=0))  # [B, K, num_head, seq_len, esz]
                        item = item[range(batch_size), selected_idx, ...]  # [B, num_head, seq_len, esz]
                        items += (item,)
                    new_key_values += (items,)
                next_past_key_values = new_key_values
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241

            logit_for_next_step = torch.stack(torch.split(logits, top_k))[range(batch_size), selected_idx, :]

            # Rebuilds the relevant parts of the model output for the selected token, for use in the next iteration
            if self.config.is_encoder_decoder:
                next_step_cross_attentions = ()
                next_step_decoder_attentions = ()
                if output_attentions:
                    for layer in outputs.cross_attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_cross_attentions += (layer,)
                    for layer in outputs.decoder_attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_decoder_attentions += (layer,)
                outputs = Seq2SeqLMOutput(
                    past_key_values=next_past_key_values,
                    decoder_hidden_states=next_decoder_hidden_states,
                    decoder_attentions=next_step_decoder_attentions or None,
                    cross_attentions=next_step_cross_attentions or None,
                )
            else:
                next_step_attentions = ()
                if output_attentions:
                    for layer in outputs.attentions:
                        layer = torch.stack(torch.split(layer, top_k, dim=0))[range(batch_size), selected_idx, ...]
                        next_step_attentions += (layer,)
                outputs = CausalLMOutputWithPast(
                    past_key_values=next_past_key_values,
                    hidden_states=next_decoder_hidden_states,
                    attentions=next_step_attentions or None,
                )
            # contrastive_search main logic end

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
2242
2243
            if streamer is not None:
                streamer.put(next_tokens.cpu())
2244
2245
2246
2247
2248
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
2249
            if eos_token_id_tensor is not None:
2250
                unfinished_sequences = unfinished_sequences.mul(
2251
2252
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )
2253

2254
2255
                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
2256
2257
                    this_peer_finished = True

2258
2259
2260
2261
2262
2263
2264
            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

2265
2266
2267
        if streamer is not None:
            streamer.end()

2268
2269
2270
2271
2272
2273
2274
2275
2276
2277
2278
2279
2280
2281
2282
2283
2284
2285
2286
2287
2288
2289
2290
2291
2292
2293
2294
2295
        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return ContrastiveSearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return ContrastiveSearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def greedy_search(
        self,
        input_ids: torch.LongTensor,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
2296
        eos_token_id: Optional[Union[int, List[int]]] = None,
2297
2298
2299
2300
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
2301
        synced_gpus: bool = False,
2302
        streamer: Optional["BaseStreamer"] = None,
2303
2304
2305
2306
2307
2308
        **model_kwargs,
    ) -> Union[GreedySearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **greedy decoding** and can be
        used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

2309
2310
2311
2312
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.greedy_search`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
2313
        guide](../generation_strategies).
2314
2315
2316
2317

        </Tip>


2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.

            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
2333
2334
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
2347
2348
2349
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372
2373
2374
2375
2376
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.GreedySearchDecoderOnlyOutput`], [`~generation.GreedySearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.GreedySearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.GreedySearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")

        >>> # set pad_token_id to eos_token_id because GPT2 does not have a PAD token
2377
        >>> model.generation_config.pad_token_id = model.generation_config.eos_token_id
2378
2379
2380
2381
2382
2383
2384

        >>> input_prompt = "It might be possible to"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
2385
        ...         MinLengthLogitsProcessor(10, eos_token_id=model.generation_config.eos_token_id),
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
        ...     ]
        ... )
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])

        >>> outputs = model.greedy_search(
        ...     input_ids, logits_processor=logits_processor, stopping_criteria=stopping_criteria
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ["It might be possible to get a better understanding of the nature of the problem, but it's not"]
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList([MaxLengthCriteria(max_length=max_length)])` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
2407
2408
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
2409
2410
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
2411
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
2412
2413
2414
2415
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
2416
        output_hidden_states = (
2417
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
2418
2419
        )
        return_dict_in_generate = (
2420
2421
2422
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
2439
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
2467
2468
2469
2470
2471
2472
2473
2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # prepare model inputs
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            # forward pass to get next token
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # pre-process distribution
            next_tokens_scores = logits_processor(input_ids, next_token_logits)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_tokens_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # argmax
            next_tokens = torch.argmax(next_tokens_scores, dim=-1)

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
2501
2502
            if streamer is not None:
                streamer.put(next_tokens.cpu())
2503
2504
2505
2506
2507
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
2508
            if eos_token_id_tensor is not None:
2509
                unfinished_sequences = unfinished_sequences.mul(
2510
2511
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )
2512

2513
2514
                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
2515
2516
                    this_peer_finished = True

2517
2518
2519
2520
2521
2522
2523
            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

2524
2525
2526
        if streamer is not None:
            streamer.end()

2527
2528
2529
2530
2531
2532
2533
2534
2535
2536
2537
2538
2539
2540
2541
2542
2543
2544
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return GreedySearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return GreedySearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def sample(
        self,
        input_ids: torch.LongTensor,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
2556
        eos_token_id: Optional[Union[int, List[int]]] = None,
2557
2558
2559
2560
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
2561
        synced_gpus: bool = False,
2562
        streamer: Optional["BaseStreamer"] = None,
2563
2564
2565
2566
2567
2568
        **model_kwargs,
    ) -> Union[SampleOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **multinomial sampling** and
        can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

2569
2570
2571
2572
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.sample`] directly. Use generate() instead.
        For an overview of generation strategies and code examples, check the [following
2573
        guide](../generation_strategies).
2574
2575
2576

        </Tip>

2577
2578
2579
2580
2581
2582
2583
2584
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
2595
2596
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
2609
2610
2611
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
2612
2613
2614
2615
2616
2617
2618
2619
2620
2621
2622
2623
2624
2625
2626
2627
2628
2629
2630
2631
2632
2633
2634
2635
2636
2637
2638
2639
2640
2641
2642
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.SampleDecoderOnlyOutput`], [`~generation.SampleEncoderDecoderOutput`] or `torch.LongTensor`:
            A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.SampleDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.SampleEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     TopKLogitsWarper,
        ...     TemperatureLogitsWarper,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")

        >>> # set pad_token_id to eos_token_id because GPT2 does not have a EOS token
        >>> model.config.pad_token_id = model.config.eos_token_id
Arthur's avatar
Arthur committed
2643
        >>> model.generation_config.pad_token_id = model.config.eos_token_id
2644
2645
2646
2647
2648
2649
2650

        >>> input_prompt = "Today is a beautiful day, and"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
2651
        ...         MinLengthLogitsProcessor(15, eos_token_id=model.generation_config.eos_token_id),
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
        ...     ]
        ... )
        >>> # instantiate logits processors
        >>> logits_warper = LogitsProcessorList(
        ...     [
        ...         TopKLogitsWarper(50),
        ...         TemperatureLogitsWarper(0.7),
        ...     ]
        ... )

        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])

        >>> torch.manual_seed(0)  # doctest: +IGNORE_RESULT
        >>> outputs = model.sample(
        ...     input_ids,
        ...     logits_processor=logits_processor,
        ...     logits_warper=logits_warper,
        ...     stopping_criteria=stopping_criteria,
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
2673
        ['Today is a beautiful day, and we must do everything possible to make it a day of celebration.']
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
2684
2685
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
2686
2687
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
2688
2689
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
2690
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
2691
2692
2693
2694
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
2695
        output_hidden_states = (
2696
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
2697
2698
        )
        return_dict_in_generate = (
2699
2700
2701
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
2718
        unfinished_sequences = torch.ones(input_ids.shape[0], dtype=torch.long, device=input_ids.device)
2719
2720
2721
2722
2723
2724
2725
2726
2727
2728
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742
2743
2744
2745
2746
2747
2748
2749
2750
2751
2752
2753
2754
2755
2756
2757
2758
2759
2760
2761
2762
2763
2764
2765
2766
2767
2768
2769
2770
2771
2772
2773
2774
2775
2776
2777
2778
2779
2780
2781
2782

        this_peer_finished = False  # used by synced_gpus only
        # auto-regressive generation
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # prepare model inputs
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            # forward pass to get next token
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # pre-process distribution
            next_token_scores = logits_processor(input_ids, next_token_logits)
            next_token_scores = logits_warper(input_ids, next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # sample
            probs = nn.functional.softmax(next_token_scores, dim=-1)
            next_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)

            # finished sentences should have their next token be a padding token
            if eos_token_id is not None:
                if pad_token_id is None:
                    raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
                next_tokens = next_tokens * unfinished_sequences + pad_token_id * (1 - unfinished_sequences)

            # update generated ids, model inputs, and length for next step
            input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
2783
2784
            if streamer is not None:
                streamer.put(next_tokens.cpu())
2785
2786
2787
2788
2789
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
2790
            if eos_token_id_tensor is not None:
2791
                unfinished_sequences = unfinished_sequences.mul(
2792
2793
                    next_tokens.tile(eos_token_id_tensor.shape[0], 1).ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0)
                )
2794

2795
2796
                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
2797
2798
                    this_peer_finished = True

2799
2800
2801
2802
2803
2804
2805
            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

2806
2807
2808
        if streamer is not None:
            streamer.end()

2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return SampleEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return SampleDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids

    def beam_search(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
2838
        eos_token_id: Optional[Union[int, List[int]]] = None,
2839
2840
2841
2842
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
2843
        synced_gpus: bool = False,
2844
2845
2846
2847
2848
2849
        **model_kwargs,
    ) -> Union[BeamSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **beam search decoding** and
        can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

2850
2851
2852
2853
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.beam_search`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
2854
        guide](../generation_strategies).
2855
2856
2857

        </Tip>

2858
2859
2860
2861
2862
2863
2864
2865
2866
2867
2868
2869
2870
2871
2872
2873
2874
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                An derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
2875
2876
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
2877
2878
2879
2880
2881
2882
2883
2884
2885
2886
2887
2888
2889
2890
2891
2892
2893
2894
2895
2896
2897
2898
2899
2900
2901
2902
2903
2904
2905
2906
2907
2908
2909
2910
2911
2912
2913
2914
2915
2916
2917
2918
2919
2920
2921
2922
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935
2936
2937
2938
2939
2940
2941
2942
2943
2944
2945
2946
2947
2948
2949
2950
2951
2952
2953
2954
2955
2956
2957
2958
2959
2960
2961
2962
2963
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.


        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs)

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        if len(stopping_criteria) == 0:
            warnings.warn("You don't have defined any stopping_criteria, this will likely loop forever", UserWarning)
2964
2965
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
2966
2967
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
2968
2969
2970
2971
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
2972
        output_hidden_states = (
2973
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
2974
2975
        )
        return_dict_in_generate = (
2976
2977
2978
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
2979
2980
2981
2982
2983
2984
2985
2986
2987
2988
2989
2990
2991
2992
2993
2994
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036
3037
3038
3039
3040
3041
3042
3043
3044
3045
3046
3047
3048
3049
3050
3051
3052
3053
3054
3055
3056
3057
3058
3059
3060
3061
3062
3063
3064
3065
3066
3067
3068
3069
3070
3071
3072
3073
3074
3075
        )

        batch_size = len(beam_scorer._beam_hyps)
        num_beams = beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam with 0 and the rest with -1e9. This makes sure that only tokens
        # of the first beam are considered to avoid sampling the exact same tokens across all beams.
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores[:, 1:] = -1e9
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]
            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)
            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores_processed,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
            next_token_scores, next_tokens = torch.topk(
                next_token_scores, 2 * num_beams, dim=1, largest=True, sorted=True
            )

3076
            next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
3077
3078
3079
3080
3081
3082
3083
3084
3085
3086
3087
3088
3089
3090
3091
3092
3093
3094
3095
3096
3097
3098
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                beam_indices=beam_indices,
            )

            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
3099
3100
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)
3101
3102
3103
3104
3105
3106
3107
3108
3109
3110
3111
3112
3113
3114
3115
3116
3117
3118
3119
3120
3121
3122
3123
3124
3125
3126
3127
3128
3129
3130
3131
3132
3133
3134
3135
3136
3137
3138
3139
3140
3141
3142
3143
3144
3145
3146
3147
3148
3149
3150
3151
3152
3153
3154
3155
3156
3157
3158
3159
3160
3161

            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def beam_sample(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        logits_warper: Optional[LogitsProcessorList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
3162
        eos_token_id: Optional[Union[int, List[int]]] = None,
3163
3164
3165
3166
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
3167
        synced_gpus: bool = False,
3168
3169
3170
3171
3172
3173
        **model_kwargs,
    ) -> Union[BeamSampleOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **beam search multinomial
        sampling** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

3174
3175
3176
3177
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.beam_sample`] directly. Use generate()
        instead. For an overview of generation strategies and code examples, check the [following
3178
        guide](../generation_strategies).
3179
3180
3181

        </Tip>

3182
3183
3184
3185
3186
3187
3188
3189
3190
3191
3192
3193
3194
3195
3196
3197
3198
3199
3200
3201
3202
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                A derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
3203
3204
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
3205
3206
3207
3208
3209
3210
3211
3212
3213
3214
3215
3216
3217
3218
3219
3220
3221
3222
3223
3224
3225
3226
3227
3228
3229
3230
3231
3232
3233
3234
3235
3236
3237
3238
3239
3240
3241
3242
3243
3244
3245
3246
3247
3248
3249
3250
3251
3252
3253
3254
3255
3256
3257
3258
3259
3260
3261
3262
3263
3264
3265
3266
3267
3268
3269
3270
3271
3272
3273
3274
3275
3276
3277
3278
3279
3280
3281
3282
3283
3284
3285
3286
3287
3288
3289
3290
3291
3292
3293
3294
3295
3296
3297
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.BeamSampleDecoderOnlyOutput`], [`~generation.BeamSampleEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSampleDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSampleEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     TopKLogitsWarper,
        ...     TemperatureLogitsWarper,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids

        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     max_length=model.config.max_length,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id)]
        ... )
        >>> # instantiate logits processors
        >>> logits_warper = LogitsProcessorList(
        ...     [
        ...         TopKLogitsWarper(50),
        ...         TemperatureLogitsWarper(0.7),
        ...     ]
        ... )

        >>> outputs = model.beam_sample(
        ...     input_ids, beam_scorer, logits_processor=logits_processor, logits_warper=logits_warper, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
3298
3299
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
3300
3301
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
3302
3303
3304
3305
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
3306
        output_hidden_states = (
3307
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
3308
3309
        )
        return_dict_in_generate = (
3310
3311
3312
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
3313
3314
3315
3316
3317
3318
3319
3320
3321
3322
3323
3324
3325
3326
3327
3328
3329
3330
3331
3332
3333
3334
3335
3336
3337
3338
3339
3340
3341
3342
3343
3344
3345
3346
3347
3348
3349
3350
3351
3352
3353
3354
3355
3356
3357
3358
3359
3360
3361
3362
3363
3364
3365
3366
3367
3368
3369
3370
3371
3372
3373
3374
        )

        batch_size = len(beam_scorer._beam_hyps)
        num_beams = beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]

            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)
            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)
3375
3376
3377
            # Note: logits warpers are intentionally applied after adding running beam scores. On some logits warpers
            # (like top_p) this is indiferent, but on others (like temperature) it is not. For reference, see
            # https://github.com/huggingface/transformers/pull/5420#discussion_r449779867
3378
3379
3380
3381
3382
3383
3384
3385
3386
3387
3388
3389
3390
3391
3392
3393
3394
3395
3396
3397
3398
3399
3400
3401
3402
3403
3404
3405
3406
3407
3408
3409
            next_token_scores = logits_warper(input_ids, next_token_scores)

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (logits_warper(input_ids, next_token_scores_processed),)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            probs = nn.functional.softmax(next_token_scores, dim=-1)

            next_tokens = torch.multinomial(probs, num_samples=2 * num_beams)
            next_token_scores = torch.gather(next_token_scores, -1, next_tokens)

            next_token_scores, _indices = torch.sort(next_token_scores, descending=True, dim=1)
            next_tokens = torch.gather(next_tokens, -1, _indices)

3410
            next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
3411
3412
3413
3414
3415
3416
3417
3418
3419
3420
3421
3422
3423
3424
3425
3426
3427
3428
3429
3430
3431
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
                beam_indices=beam_indices,
            )
            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
3432
3433
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)
3434
3435
3436
3437
3438
3439
3440
3441
3442
3443
3444
3445
3446
3447
3448
3449
3450
3451
3452
3453
3454
3455
3456
3457
3458
3459
3460
3461
3462
3463
3464
3465
3466
3467
3468
3469
3470
3471
3472
3473
3474
3475
3476
3477
3478
3479
3480
3481
3482
3483
3484
3485
3486
3487
3488
3489
3490
3491
3492
3493

            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSampleEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSampleDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def group_beam_search(
        self,
        input_ids: torch.LongTensor,
        beam_scorer: BeamScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
3494
        eos_token_id: Optional[Union[int, List[int]]] = None,
3495
3496
3497
3498
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
3499
        synced_gpus: bool = False,
3500
3501
3502
3503
3504
3505
        **model_kwargs,
    ):
        r"""
        Generates sequences of token ids for models with a language modeling head using **diverse beam search
        decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

3506
3507
3508
3509
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.group_beam_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
3510
        guide](../generation_strategies).
3511
3512
3513

        </Tip>

3514
3515
3516
3517
3518
3519
3520
3521
3522
3523
3524
3525
3526
3527
3528
3529
3530
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            beam_scorer (`BeamScorer`):
                An derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation. For more information, the documentation of [`BeamScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
3531
3532
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
3533
3534
3535
3536
3537
3538
3539
3540
3541
3542
3543
3544
3545
3546
3547
3548
3549
3550
3551
3552
3553
3554
3555
3556
3557
3558
3559
3560
3561
3562
3563
3564
3565
3566
3567
3568
3569
3570
3571
3572
3573
3574
3575
3576
3577
3578
3579
3580
3581
3582
3583
3584
3585
3586
3587
3588
3589
3590
3591
3592
3593
3594
3595
3596
3597
3598
3599
3600
3601
3602
3603
3604
3605
3606
3607
3608
3609
3610
3611
3612
3613
3614
3615
3616
3617
3618
3619
3620
3621
3622
3623
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)

            model_kwargs:
                Additional model specific kwargs that will be forwarded to the `forward` function of the model. If
                model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if [`~generation.BeamSearchDecoderOnlyOutput`] if
            `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a
            [`~generation.BeamSearchEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     HammingDiversityLogitsProcessor,
        ...     BeamSearchScorer,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run diverse beam search using 6 beams
        >>> num_beams = 6
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> # instantiate beam scorer
        >>> beam_scorer = BeamSearchScorer(
        ...     batch_size=1,
        ...     max_length=model.config.max_length,
        ...     num_beams=num_beams,
        ...     device=model.device,
        ...     num_beam_groups=3,
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         HammingDiversityLogitsProcessor(5.5, num_beams=6, num_beam_groups=3),
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.group_beam_search(
        ...     input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt bist du?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
3624
3625
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
3626
3627
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
3628
3629
3630
3631
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
3632
        output_hidden_states = (
3633
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
3634
3635
        )
        return_dict_in_generate = (
3636
3637
3638
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
3639
3640
3641
3642
3643
        )

        num_beams = beam_scorer.num_beams
        num_beam_groups = beam_scorer.num_beam_groups
        num_sub_beams = num_beams // num_beam_groups
3644
        batch_size = len(beam_scorer._beam_hyps) // num_beam_groups
3645
3646
3647
3648
3649
3650
3651
3652
3653
3654
3655
3656
3657
3658
3659
3660
3661
3662
3663
3664
3665
3666
3667
3668
3669
3670
3671
3672
3673
3674
3675
3676
3677
3678
3679
3680
3681
3682
3683
3684
3685
3686
3687
3688
3689
3690
3691
3692
3693
3694
3695
3696
3697
3698
3699
3700
3701
3702
3703
3704
3705
3706
3707
3708
3709
3710
3711
3712
3713
3714
3715
3716
3717
3718
3719
3720
3721
3722
3723
3724
3725
3726
3727
3728
3729
3730
3731
3732
3733
3734
3735
3736
3737
3738
3739
3740
3741
3742
3743
3744
3745
3746
3747
3748
3749
3750
3751
3752
3753
        device = input_ids.device

        batch_beam_size, cur_len = input_ids.shape

        if return_dict_in_generate and output_scores:
            beam_indices = [tuple(() for _ in range(num_sub_beams * batch_size)) for _ in range(num_beam_groups)]
        else:
            beam_indices = None

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam of each group with 0 and the rest with -1e9. This ensures that the beams in
        # the same group don't produce same tokens everytime.
        beam_scores = torch.full((batch_size, num_beams), -1e9, dtype=torch.float, device=device)
        beam_scores[:, ::num_sub_beams] = 0
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # predicted tokens in cur_len step
            current_tokens = torch.zeros(batch_size * num_beams, dtype=input_ids.dtype, device=device)

            # indices which will form the beams in the next time step
            reordering_indices = torch.zeros(batch_size * num_beams, dtype=torch.long, device=device)

            # do one decoder step on all beams of all sentences in batch
            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            if output_scores:
                processed_score = torch.zeros_like(outputs.logits[:, -1, :])

            for beam_group_idx in range(num_beam_groups):
                group_start_idx = beam_group_idx * num_sub_beams
                group_end_idx = min(group_start_idx + num_sub_beams, num_beams)
                group_size = group_end_idx - group_start_idx

                # indices of beams of current group among all sentences in batch
                batch_group_indices = []

                for batch_idx in range(batch_size):
                    batch_group_indices.extend(
                        [batch_idx * num_beams + idx for idx in range(group_start_idx, group_end_idx)]
                    )
                group_input_ids = input_ids[batch_group_indices]

                # select outputs of beams of current group only
                next_token_logits = outputs.logits[batch_group_indices, -1, :]

                # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
                # cannot be generated both before and after the `nn.functional.log_softmax` operation.
                next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
                next_token_scores = nn.functional.log_softmax(
                    next_token_logits, dim=-1
                )  # (batch_size * group_size, vocab_size)
                vocab_size = next_token_scores.shape[-1]

                next_token_scores_processed = logits_processor(
                    group_input_ids, next_token_scores, current_tokens=current_tokens, beam_group_idx=beam_group_idx
                )
                next_token_scores = next_token_scores_processed + beam_scores[batch_group_indices].unsqueeze(-1)
                next_token_scores = next_token_scores.expand_as(next_token_scores_processed)

                if output_scores:
                    processed_score[batch_group_indices] = next_token_scores_processed

                # reshape for beam search
                next_token_scores = next_token_scores.view(batch_size, group_size * vocab_size)

                # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
                next_token_scores, next_tokens = torch.topk(
                    next_token_scores, 2 * group_size, dim=1, largest=True, sorted=True
                )

3754
                next_indices = torch.div(next_tokens, vocab_size, rounding_mode="floor")
3755
3756
3757
3758
3759
3760
3761
3762
3763
3764
3765
3766
                next_tokens = next_tokens % vocab_size

                # stateless
                process_beam_indices = sum(beam_indices, ()) if beam_indices is not None else None
                beam_outputs = beam_scorer.process(
                    group_input_ids,
                    next_token_scores,
                    next_tokens,
                    next_indices,
                    pad_token_id=pad_token_id,
                    eos_token_id=eos_token_id,
                    beam_indices=process_beam_indices,
3767
                    group_index=beam_group_idx,
3768
3769
3770
3771
3772
3773
3774
3775
3776
3777
3778
3779
3780
3781
3782
3783
3784
                )
                beam_scores[batch_group_indices] = beam_outputs["next_beam_scores"]
                beam_next_tokens = beam_outputs["next_beam_tokens"]
                beam_idx = beam_outputs["next_beam_indices"]

                if return_dict_in_generate and output_scores:
                    beam_indices[beam_group_idx] = tuple(
                        beam_indices[beam_group_idx][beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices[0]))
                    )

                input_ids[batch_group_indices] = group_input_ids[beam_idx]
                group_input_ids = torch.cat([group_input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)
                current_tokens[batch_group_indices] = group_input_ids[:, -1]

                # (beam_idx // group_size) -> batch_idx
                # (beam_idx % group_size) -> offset of idx inside the group
                reordering_indices[batch_group_indices] = (
3785
3786
3787
                    num_beams * torch.div(beam_idx, group_size, rounding_mode="floor")
                    + group_start_idx
                    + (beam_idx % group_size)
3788
3789
3790
3791
3792
3793
3794
3795
3796
3797
3798
3799
3800
3801
3802
3803
3804
3805
3806
3807
3808
3809
3810
3811
3812
                )

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (processed_score,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            input_ids = torch.cat([input_ids, current_tokens.unsqueeze(-1)], dim=-1)

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
3813
3814
3815
3816
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(
                    model_kwargs["past_key_values"], reordering_indices
                )
3817
3818
3819
3820
3821
3822
3823
3824
3825
3826
3827
3828
3829
3830
3831
3832
3833
3834
3835
3836
3837
3838
3839
3840
3841
3842
3843
3844
3845
3846
3847
3848
3849
3850
3851
3852
3853
3854
3855
3856
3857
3858
3859
3860
3861
3862
3863
3864
3865
3866
3867
3868
3869
3870
3871
3872
3873
3874

            # increase cur_len
            cur_len = cur_len + 1

            if beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        final_beam_indices = sum(beam_indices, ()) if beam_indices is not None else None
        sequence_outputs = beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
            beam_indices=final_beam_indices,
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None

            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
                    beam_indices=sequence_outputs["beam_indices"],
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

    def constrained_beam_search(
        self,
        input_ids: torch.LongTensor,
        constrained_beam_scorer: ConstrainedBeamSearchScorer,
        logits_processor: Optional[LogitsProcessorList] = None,
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        max_length: Optional[int] = None,
        pad_token_id: Optional[int] = None,
3875
        eos_token_id: Optional[Union[int, List[int]]] = None,
3876
3877
3878
3879
3880
3881
3882
3883
3884
3885
3886
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: Optional[bool] = None,
        **model_kwargs,
    ) -> Union[BeamSearchOutput, torch.LongTensor]:
        r"""
        Generates sequences of token ids for models with a language modeling head using **constrained beam search
        decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models.

3887
3888
3889
3890
        <Tip warning={true}>

        In most cases, you do not need to call [`~generation.GenerationMixin.constrained_beam_search`] directly. Use
        generate() instead. For an overview of generation strategies and code examples, check the [following
3891
        guide](../generation_strategies).
3892
3893
3894

        </Tip>

3895
3896
3897
3898
3899
3900
3901
3902
3903
3904
3905
3906
3907
3908
3909
3910
3911
3912
3913
3914
3915
3916
        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            constrained_beam_scorer (`ConstrainedBeamSearchScorer`):
                A derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and
                sorted during generation, while satisfying a list of positive constraints. For more information, the
                documentation of [`ConstrainedBeamSearchScorer`] should be read.
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
            max_length (`int`, *optional*, defaults to 20):
                **DEPRECATED**. Use `logits_processor` or `stopping_criteria` directly to cap the number of generated
                tokens. The maximum length of the sequence to be generated.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
3917
3918
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
3919
3920
3921
3922
3923
3924
3925
3926
3927
3928
3929
3930
3931
3932
3933
3934
3935
3936
3937
3938
3939
3940
3941
3942
3943
3944
3945
3946
3947
3948
3949
3950
3951
3952
3953
3954
3955
3956
3957
3958
3959
3960
3961
3962
3963
3964
3965
3966
3967
3968
3969
3970
3971
3972
3973
3974
3975
3976
3977
3978
3979
3980
3981
3982
3983
3984
3985
3986
3987
3988
3989
3990
3991
3992
3993
3994
3995
3996
3997
3998
3999
4000
4001
4002
4003
4004
4005
4006
4007
4008
4009
4010
4011
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            model_kwargs:
                Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is
                an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`generation.BeamSearchDecoderOnlyOutput`], [`~generation.BeamSearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.BeamSearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.BeamSearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.


        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForSeq2SeqLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     ConstrainedBeamSearchScorer,
        ...     PhrasalConstraint,
        ... )
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("t5-base")
        >>> model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        >>> encoder_input_str = "translate English to German: How old are you?"
        >>> encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids


        >>> # lets run beam search using 3 beams
        >>> num_beams = 3
        >>> # define decoder start token ids
        >>> input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
        >>> input_ids = input_ids * model.config.decoder_start_token_id

        >>> # add encoder_outputs to model keyword arguments
        >>> model_kwargs = {
        ...     "encoder_outputs": model.get_encoder()(
        ...         encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
        ...     )
        ... }

        >>> constraint_str = "Sie"
        >>> constraint_token_ids = tokenizer.encode(constraint_str)[:-1]  # slice to remove eos token
        >>> constraints = [PhrasalConstraint(token_ids=constraint_token_ids)]


        >>> # instantiate beam scorer
        >>> beam_scorer = ConstrainedBeamSearchScorer(
        ...     batch_size=1, num_beams=num_beams, device=model.device, constraints=constraints
        ... )

        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
        ...     ]
        ... )

        >>> outputs = model.constrained_beam_search(
        ...     input_ids, beam_scorer, constraints=constraints, logits_processor=logits_processor, **model_kwargs
        ... )

        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ['Wie alt sind Sie?']
        ```"""
        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        if max_length is not None:
            warnings.warn(
                "`max_length` is deprecated in this function, use"
                " `stopping_criteria=StoppingCriteriaList(MaxLengthCriteria(max_length=max_length))` instead.",
                UserWarning,
            )
            stopping_criteria = validate_stopping_criteria(stopping_criteria, max_length)
        if len(stopping_criteria) == 0:
            warnings.warn("You don't have defined any stopping_criteria, this will likely loop forever", UserWarning)
4012
4013
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
4014
4015
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
4016
4017
4018
4019
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
4020
        output_hidden_states = (
4021
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
4022
4023
        )
        return_dict_in_generate = (
4024
4025
4026
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
4027
4028
        )

4029
4030
4031
4032
4033
4034
4035
4036
4037
4038
        batch_size = len(constrained_beam_scorer._beam_hyps)
        num_beams = constrained_beam_scorer.num_beams

        batch_beam_size, cur_len = input_ids.shape

        if num_beams * batch_size != batch_beam_size:
            raise ValueError(
                f"Batch dimension of `input_ids` should be {num_beams * batch_size}, but is {batch_beam_size}."
            )

4039
4040
        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
4041
4042
4043
        beam_indices = (
            tuple(() for _ in range(batch_beam_size)) if (return_dict_in_generate and output_scores) else None
        )
4044
4045
4046
4047
4048
4049
4050
4051
4052
4053
4054
4055
4056
4057
4058
4059
4060
4061
4062
4063
4064
4065
4066
4067
4068
4069
4070
4071
4072
4073
4074
4075
4076
4077
4078
4079
4080
4081
4082
4083
4084
4085
4086
4087
4088
4089
4090
4091
4092
4093
4094
4095
4096
4097
4098
4099
4100
4101
4102
4103
4104
4105
4106
4107
4108
4109
4110
4111
4112
4113
4114
4115
4116
4117
4118
4119
4120
4121
4122
4123
4124
4125
4126
4127
4128
4129
4130
4131
4132
4133
4134
4135
4136
4137
4138
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # initialise score of first beam with 0 and the rest with -1e9. This makes sure that only tokens
        # of the first beam are considered to avoid sampling the exact same tokens across all beams.
        beam_scores = torch.zeros((batch_size, num_beams), dtype=torch.float, device=input_ids.device)
        beam_scores[:, 1:] = -1e9
        beam_scores = beam_scores.view((batch_size * num_beams,))

        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)

            outputs = self(
                **model_inputs,
                return_dict=True,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )

            if synced_gpus and this_peer_finished:
                cur_len = cur_len + 1
                continue  # don't waste resources running the code we don't need

            next_token_logits = outputs.logits[:, -1, :]
            # hack: adjust tokens for Marian. For Marian we have to make sure that the `pad_token_id`
            # cannot be generated both before and after the `nn.functional.log_softmax` operation.
            next_token_logits = self.adjust_logits_during_generation(next_token_logits, cur_len=cur_len)
            next_token_scores = nn.functional.log_softmax(
                next_token_logits, dim=-1
            )  # (batch_size * num_beams, vocab_size)

            next_token_scores_processed = logits_processor(input_ids, next_token_scores)

            next_token_scores = next_token_scores_processed + beam_scores[:, None].expand_as(next_token_scores)

            scores_for_all_vocab = next_token_scores.clone()

            # Store scores, attentions and hidden_states when required
            if return_dict_in_generate:
                if output_scores:
                    scores += (next_token_scores,)
                if output_attentions:
                    decoder_attentions += (
                        (outputs.decoder_attentions,) if self.config.is_encoder_decoder else (outputs.attentions,)
                    )
                    if self.config.is_encoder_decoder:
                        cross_attentions += (outputs.cross_attentions,)

                if output_hidden_states:
                    decoder_hidden_states += (
                        (outputs.decoder_hidden_states,)
                        if self.config.is_encoder_decoder
                        else (outputs.hidden_states,)
                    )

            # reshape for beam search
            vocab_size = next_token_scores.shape[-1]
            next_token_scores = next_token_scores.view(batch_size, num_beams * vocab_size)

            # Sample 2 next tokens for each beam (so we have some spare tokens and match output of beam search)
            next_token_scores, next_tokens = torch.topk(
                next_token_scores, 2 * num_beams, dim=1, largest=True, sorted=True
            )

            next_indices = (next_tokens / vocab_size).long()
            next_tokens = next_tokens % vocab_size

            # stateless
            beam_outputs = constrained_beam_scorer.process(
                input_ids,
                next_token_scores,
                next_tokens,
                next_indices,
                scores_for_all_vocab,
                pad_token_id=pad_token_id,
                eos_token_id=eos_token_id,
4139
                beam_indices=beam_indices,
4140
4141
4142
4143
4144
4145
4146
4147
4148
            )
            beam_scores = beam_outputs["next_beam_scores"]
            beam_next_tokens = beam_outputs["next_beam_tokens"]
            beam_idx = beam_outputs["next_beam_indices"]

            input_ids = torch.cat([input_ids[beam_idx, :], beam_next_tokens.unsqueeze(-1)], dim=-1)
            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )
4149
4150
            if model_kwargs["past_key_values"] is not None:
                model_kwargs["past_key_values"] = self._reorder_cache(model_kwargs["past_key_values"], beam_idx)
4151

4152
4153
4154
            if return_dict_in_generate and output_scores:
                beam_indices = tuple((beam_indices[beam_idx[i]] + (beam_idx[i],) for i in range(len(beam_indices))))

4155
4156
4157
4158
4159
4160
4161
4162
4163
4164
4165
4166
4167
4168
4169
4170
4171
            # increase cur_len
            cur_len = cur_len + 1

            if constrained_beam_scorer.is_done or stopping_criteria(input_ids, scores):
                if not synced_gpus:
                    break
                else:
                    this_peer_finished = True

        sequence_outputs = constrained_beam_scorer.finalize(
            input_ids,
            beam_scores,
            next_tokens,
            next_indices,
            pad_token_id=pad_token_id,
            eos_token_id=eos_token_id,
            max_length=stopping_criteria.max_length,
4172
            beam_indices=beam_indices,
4173
4174
4175
4176
4177
4178
4179
4180
4181
4182
        )

        if return_dict_in_generate:
            if not output_scores:
                sequence_outputs["sequence_scores"] = None
            if self.config.is_encoder_decoder:
                return BeamSearchEncoderDecoderOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
4183
                    beam_indices=sequence_outputs["beam_indices"],
4184
4185
4186
4187
4188
4189
4190
4191
4192
4193
4194
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return BeamSearchDecoderOnlyOutput(
                    sequences=sequence_outputs["sequences"],
                    sequences_scores=sequence_outputs["sequence_scores"],
                    scores=scores,
4195
                    beam_indices=sequence_outputs["beam_indices"],
4196
4197
4198
4199
4200
4201
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return sequence_outputs["sequences"]

4202
    def assisted_decoding(
4203
4204
4205
        self,
        input_ids: torch.LongTensor,
        assistant_model: "PreTrainedModel",
4206
        do_sample: bool = False,
4207
        logits_processor: Optional[LogitsProcessorList] = None,
4208
        logits_warper: Optional[LogitsProcessorList] = None,
4209
4210
4211
4212
4213
4214
4215
4216
4217
4218
4219
4220
        stopping_criteria: Optional[StoppingCriteriaList] = None,
        pad_token_id: Optional[int] = None,
        eos_token_id: Optional[Union[int, List[int]]] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        output_scores: Optional[bool] = None,
        return_dict_in_generate: Optional[bool] = None,
        synced_gpus: bool = False,
        streamer: Optional["BaseStreamer"] = None,
        **model_kwargs,
    ):
        r"""
4221
4222
4223
        Generates sequences of token ids for models with a language modeling head using **greedy decoding** or
        **sample** (depending on `do_sample`), assisted by a smaller model. Can be used for text-decoder, text-to-text,
        speech-to-text, and vision-to-text models.
4224
4225
4226

        <Tip warning={true}>

4227
        In most cases, you do not need to call [`~generation.GenerationMixin.assisted_decoding`] directly. Use
4228
4229
4230
4231
4232
4233
4234
4235
4236
4237
4238
4239
4240
        generate() instead. For an overview of generation strategies and code examples, check the [following
        guide](../generation_strategies).

        </Tip>

        Parameters:
            input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
                The sequence used as a prompt for the generation.
            assistant_model (`PreTrainedModel`, *optional*):
                An assistant model that can be used to accelerate generation. The assistant model must have the exact
                same tokenizer. The acceleration is achieved when forecasting candidate tokens with the assistent model
                is much faster than running generation with the model you're calling generate from. As such, the
                assistant model should be much smaller.
4241
4242
            do_sample (`bool`, *optional*, defaults to `False`):
                Whether or not to use sampling ; use greedy decoding otherwise.
4243
4244
4245
            logits_processor (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`]
                used to modify the prediction scores of the language modeling head applied at each generation step.
4246
4247
4248
4249
            logits_warper (`LogitsProcessorList`, *optional*):
                An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsWarper`] used
                to warp the prediction score distribution of the language modeling head applied before multinomial
                sampling at each generation step.
4250
4251
4252
4253
4254
4255
4256
4257
4258
4259
4260
4261
4262
4263
4264
4265
4266
4267
4268
4269
4270
4271
4272
4273
4274
4275
4276
4277
4278
4279
4280
4281
4282
4283
4284
4285
4286
4287
4288
4289
4290
4291
4292
4293
4294
4295
4296
4297
4298
4299
4300
4301
4302
4303
4304
4305
4306
4307
4308
            stopping_criteria (`StoppingCriteriaList`, *optional*):
                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
                used to tell if the generation loop should stop.
            pad_token_id (`int`, *optional*):
                The id of the *padding* token.
            eos_token_id (`Union[int, List[int]]`, *optional*):
                The id of the *end-of-sequence* token. Optionally, use a list to set multiple *end-of-sequence* tokens.
            output_attentions (`bool`, *optional*, defaults to `False`):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more details.
            output_hidden_states (`bool`, *optional*, defaults to `False`):
                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
                for more details.
            output_scores (`bool`, *optional*, defaults to `False`):
                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
            synced_gpus (`bool`, *optional*, defaults to `False`):
                Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
            streamer (`BaseStreamer`, *optional*):
                Streamer object that will be used to stream the generated sequences. Generated tokens are passed
                through `streamer.put(token_ids)` and the streamer is responsible for any further processing.
            model_kwargs:
                Additional model specific keyword arguments will be forwarded to the `forward` function of the model.
                If model is an encoder-decoder model the kwargs should include `encoder_outputs`.

        Return:
            [`~generation.GreedySearchDecoderOnlyOutput`], [`~generation.GreedySearchEncoderDecoderOutput`] or
            `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a
            [`~generation.GreedySearchDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and
            `return_dict_in_generate=True` or a [`~generation.GreedySearchEncoderDecoderOutput`] if
            `model.config.is_encoder_decoder=True`.

        Examples:

        ```python
        >>> from transformers import (
        ...     AutoTokenizer,
        ...     AutoModelForCausalLM,
        ...     LogitsProcessorList,
        ...     MinLengthLogitsProcessor,
        ...     StoppingCriteriaList,
        ...     MaxLengthCriteria,
        ... )

        >>> tokenizer = AutoTokenizer.from_pretrained("gpt2")
        >>> model = AutoModelForCausalLM.from_pretrained("gpt2")
        >>> assistant_model = AutoModelForCausalLM.from_pretrained("distilgpt2")
        >>> # set pad_token_id to eos_token_id because GPT2 does not have a PAD token
        >>> model.generation_config.pad_token_id = model.generation_config.eos_token_id
        >>> input_prompt = "It might be possible to"
        >>> input_ids = tokenizer(input_prompt, return_tensors="pt").input_ids
        >>> # instantiate logits processors
        >>> logits_processor = LogitsProcessorList(
        ...     [
        ...         MinLengthLogitsProcessor(10, eos_token_id=model.generation_config.eos_token_id),
        ...     ]
        ... )
        >>> stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=20)])
4309
        >>> outputs = model.assisted_decoding(
4310
4311
4312
4313
4314
4315
4316
4317
4318
4319
4320
4321
4322
4323
        ...     input_ids,
        ...     assistant_model=assistant_model,
        ...     logits_processor=logits_processor,
        ...     stopping_criteria=stopping_criteria,
        ... )
        >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
        ["It might be possible to get a better understanding of the nature of the problem, but it's not"]
        ```"""
        # Assistant: initialize assistant-related variables
        if not hasattr(assistant_model, "max_assistant_tokens"):
            assistant_model.max_assistant_tokens = 5  # this value, which will be updated, persists across calls

        # init values
        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
4324
        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
4325
4326
4327
4328
4329
4330
4331
4332
4333
4334
4335
4336
4337
4338
4339
4340
4341
4342
4343
4344
4345
4346
4347
4348
4349
4350
4351
4352
4353
4354
4355
4356
4357
4358
4359
4360
4361
        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
        pad_token_id = pad_token_id if pad_token_id is not None else self.generation_config.pad_token_id
        eos_token_id = eos_token_id if eos_token_id is not None else self.generation_config.eos_token_id
        if eos_token_id is not None and pad_token_id is None:
            raise ValueError("If `eos_token_id` is defined, make sure that `pad_token_id` is defined.")
        if isinstance(eos_token_id, int):
            eos_token_id = [eos_token_id]
        eos_token_id_tensor = torch.tensor(eos_token_id).to(input_ids.device) if eos_token_id is not None else None
        output_scores = output_scores if output_scores is not None else self.generation_config.output_scores
        output_attentions = (
            output_attentions if output_attentions is not None else self.generation_config.output_attentions
        )
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.generation_config.output_hidden_states
        )
        return_dict_in_generate = (
            return_dict_in_generate
            if return_dict_in_generate is not None
            else self.generation_config.return_dict_in_generate
        )

        # init attention / hidden states / scores tuples
        scores = () if (return_dict_in_generate and output_scores) else None
        decoder_attentions = () if (return_dict_in_generate and output_attentions) else None
        cross_attentions = () if (return_dict_in_generate and output_attentions) else None
        decoder_hidden_states = () if (return_dict_in_generate and output_hidden_states) else None

        # if model is an encoder-decoder, retrieve encoder attention weights and hidden states
        if return_dict_in_generate and self.config.is_encoder_decoder:
            encoder_attentions = model_kwargs["encoder_outputs"].get("attentions") if output_attentions else None
            encoder_hidden_states = (
                model_kwargs["encoder_outputs"].get("hidden_states") if output_hidden_states else None
            )

        # keep track of which sequences are already finished
        unfinished_sequences = input_ids.new(input_ids.shape[0]).fill_(1)

4362
4363
        # other auxiliary variables
        max_len = stopping_criteria[0].max_length
4364
4365
4366
4367
4368
4369
4370
4371
4372
        assistant_kv_indexing = (
            1
            if "bloom" in assistant_model.__class__.__name__.lower()
            or (
                assistant_model.config.architectures is not None
                and "bloom" in assistant_model.config.architectures[0].lower()
            )
            else 0
        )
4373

4374
4375
4376
4377
4378
4379
4380
4381
4382
4383
4384
4385
4386
4387
4388
4389
4390
4391
4392
4393
4394
4395
        this_peer_finished = False  # used by synced_gpus only
        while True:
            if synced_gpus:
                # Under synced_gpus the `forward` call must continue until all gpus complete their sequence.
                # The following logic allows an early break if all peers finished generating their sequence
                this_peer_finished_flag = torch.tensor(0.0 if this_peer_finished else 1.0).to(input_ids.device)
                # send 0.0 if we finished, 1.0 otherwise
                dist.all_reduce(this_peer_finished_flag, op=dist.ReduceOp.SUM)
                # did all peers finish? the reduced sum will be 0.0 then
                if this_peer_finished_flag.item() == 0.0:
                    break

            # Assistant: main logic start
            cur_len = input_ids.shape[-1]

            #  1. Forecast next N tokens using the assistant model. This `for` block can be replaced with a
            # `.generate()` call if we decide to add `past_key_values` as a possible output of generate, as we
            # need access to the assistant cache to secure strong speedups.
            candidate_input_ids = input_ids
            for _ in range(int(assistant_model.max_assistant_tokens)):
                # 1.1. use the assistant model to obtain the next candidate logits
                if "assistant_past_key_values" in model_kwargs:
4396
                    prev_seq_len = model_kwargs["assistant_past_key_values"][0][assistant_kv_indexing].shape[-2]
4397
4398
                    # `new_token_len` can be 1 or 2 (next token in assistant + last token picked by the larger model)
                    new_token_len = candidate_input_ids.shape[1] - prev_seq_len
4399
4400
                    assist_inputs = candidate_input_ids[:, -new_token_len:]
                    assist_attn = torch.ones_like(candidate_input_ids)
4401
4402
4403
                    # TODO (joao): make it compatible with models that use unconventional fwd pass logic, like blip2
                    if assistant_model.config.is_encoder_decoder:
                        assistant_model_outputs = assistant_model(
4404
4405
                            decoder_input_ids=assist_inputs,
                            decoder_attention_mask=assist_attn,
4406
4407
4408
4409
4410
                            past_key_values=model_kwargs["assistant_past_key_values"],
                            encoder_outputs=model_kwargs["assistant_encoder_outputs"],
                        )
                    else:
                        assistant_model_outputs = assistant_model(
4411
4412
                            assist_inputs,
                            attention_mask=assist_attn,
4413
4414
4415
4416
4417
4418
4419
4420
4421
4422
4423
4424
4425
4426
4427
4428
4429
4430
4431
4432
4433
4434
4435
4436
4437
4438
4439
4440
4441
4442
4443
4444
4445
4446
                            past_key_values=model_kwargs["assistant_past_key_values"],
                        )
                else:
                    if assistant_model.config.is_encoder_decoder:
                        assistant_model_outputs = assistant_model(
                            decoder_input_ids=candidate_input_ids,
                            encoder_outputs=model_kwargs["assistant_encoder_outputs"],
                        )
                    else:
                        assistant_model_outputs = assistant_model(candidate_input_ids)

                # 1.2. greedily select the next candidate token
                model_kwargs["assistant_past_key_values"] = assistant_model_outputs.past_key_values
                if len(logits_processor) > 0:
                    assistant_model_outputs.logits[:, -1, :] = logits_processor(
                        candidate_input_ids, assistant_model_outputs.logits[:, -1, :]
                    )
                new_token = assistant_model_outputs.logits[:, -1, :].argmax(dim=-1)
                candidate_input_ids = torch.cat((candidate_input_ids, new_token[:, None]), dim=-1)

                # 1.3. stop assistant generation on EOS
                if eos_token_id_tensor is not None:
                    last_assistant_token_is_eos = new_token.tile(eos_token_id_tensor.shape[0], 1)
                    last_assistant_token_is_eos = (
                        ~last_assistant_token_is_eos.ne(eos_token_id_tensor.unsqueeze(1)).prod(dim=0).bool()
                    )
                    if last_assistant_token_is_eos:
                        break
                else:
                    last_assistant_token_is_eos = False

            candidate_length = candidate_input_ids.shape[1] - input_ids.shape[1]

            # 2. Use the original model to obtain the next token logits given the candidate sequence. We obtain
4447
4448
            # `candidate_length + 1` relevant logits from this process: in the event that all candidates are correct,
            # we use this forward pass to also pick the subsequent logits in the original model.
4449
4450

            # 2.1. Run a forward pass on the candidate sequence
4451
            if "past_key_values" in model_kwargs:
4452
4453
                model_attn = torch.ones_like(candidate_input_ids)
                model_input_ids = candidate_input_ids[:, -candidate_length - 1 :]
4454
4455
                if self.config.is_encoder_decoder:
                    outputs = self(
4456
4457
                        decoder_input_ids=model_input_ids,
                        decoder_attention_mask=model_attn,
4458
4459
4460
4461
                        past_key_values=model_kwargs["past_key_values"],
                        encoder_outputs=model_kwargs["encoder_outputs"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
4462
                        use_cache=True,
4463
4464
4465
                    )
                else:
                    outputs = self(
4466
4467
                        model_input_ids,
                        attention_mask=model_attn,
4468
4469
4470
                        past_key_values=model_kwargs["past_key_values"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
4471
                        use_cache=True,
4472
4473
4474
4475
4476
4477
4478
4479
                    )
            else:
                if self.config.is_encoder_decoder:
                    outputs = self(
                        decoder_input_ids=candidate_input_ids,
                        encoder_outputs=model_kwargs["encoder_outputs"],
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
4480
                        use_cache=True,
4481
4482
4483
4484
4485
4486
                    )
                else:
                    outputs = self(
                        candidate_input_ids,
                        output_attentions=output_attentions,
                        output_hidden_states=output_hidden_states,
4487
                        use_cache=True,
4488
4489
                    )

4490
            # 2.2. Process the new logits
4491
4492
4493
4494
            new_logits = outputs.logits[:, -candidate_length - 1 :]  # excludes the input prompt if present
            if len(logits_processor) > 0:
                for i in range(candidate_length):
                    new_logits[:, i, :] = logits_processor(candidate_input_ids[:, : cur_len + i], new_logits[:, i, :])
4495
4496
4497
4498
            if len(logits_warper) > 0:
                for i in range(candidate_length):
                    new_logits[:, i, :] = logits_warper(candidate_input_ids[:, : cur_len + i], new_logits[:, i, :])

4499
            # 3. Obtain the next tokens from the original model logits.
4500
4501
            if do_sample:
                probs = new_logits[:, -candidate_length - 1 :, :].softmax(dim=-1)
4502
                selected_tokens = torch.multinomial(probs[0, :, :], num_samples=1).squeeze(1)[None, :]
4503
            else:
4504
                selected_tokens = new_logits[:, -candidate_length - 1 :, :].argmax(dim=-1)
4505
4506
4507
4508

            # 4. Compare the argmax from the original model logits with the assistant forecasted tokens. We can keep
            # the assistant forecasted tokens until the first mismatch, or until the max length is reached.
            candidate_new_tokens = candidate_input_ids[:, -candidate_length:]
4509
            n_matches = ((~(candidate_new_tokens == selected_tokens[:, :-1])).cumsum(dim=-1) < 1).sum()
4510

4511
4512
4513
4514
            # 5. Update variables according to the number of matching assistant tokens. Remember: the token generated
            # by the model after the last candidate match is also valid, as it is generated from a correct sequence.
            # Because of this last token, assisted generation search reduces to a normal greedy search/sample if there
            # is no match.
4515

4516
            # 5.1. Ensure we don't generate beyond max_len or an EOS token
4517
4518
            if last_assistant_token_is_eos and n_matches == candidate_length:
                n_matches -= 1
4519
4520
4521
4522
4523
            n_matches = min(n_matches, max_len - cur_len - 1)

            # 5.2. Get the valid continuation, after the matching tokens
            valid_tokens = selected_tokens[:, : n_matches + 1]
            input_ids = torch.cat((input_ids, valid_tokens), dim=-1)
4524
            if streamer is not None:
4525
4526
                streamer.put(valid_tokens.cpu())
            new_cur_len = input_ids.shape[-1]
4527

4528
4529
4530
            # 5.3. Discard past key values relative to unused assistant tokens
            new_cache_size = new_cur_len - 1
            outputs.past_key_values = _crop_past_key_values(self, outputs.past_key_values, new_cache_size)
4531
            model_kwargs["assistant_past_key_values"] = _crop_past_key_values(
4532
4533
                assistant_model, model_kwargs["assistant_past_key_values"], new_cache_size - 1
            )  # the assistant does not have the token after the last match, hence the -1
4534

4535
4536
4537
4538
4539
            # 6. Adjust the max number of assistant tokens to use in the next iteration. This is a simple heuristic,
            # probably can be improved -- we want to balance the benefits of getting assistant tokens correct with the
            # cost of forecasting incorrect assistant tokens.
            if n_matches == int(assistant_model.max_assistant_tokens):
                assistant_model.max_assistant_tokens += 2.0
4540
            else:
4541
                assistant_model.max_assistant_tokens = max(1.0, assistant_model.max_assistant_tokens - 1.0)
4542

4543
            # Assistant: main logic end
4544
4545
4546
4547
4548
4549
4550
4551
4552
4553
4554

            if synced_gpus and this_peer_finished:
                continue  # don't waste resources running the code we don't need

            # Store scores, attentions and hidden_states when required
            # Assistant: modified to append one tuple element per token, as in the other generation methods.
            if return_dict_in_generate:
                if output_scores:
                    scores += tuple(new_logits[:, i, :] for i in range(n_matches + 1))

                if "past_key_values" not in model_kwargs:
4555
                    added_len = new_cur_len
4556
                else:
4557
                    added_len = n_matches + 1
4558
4559
4560
4561

                if output_attentions:
                    if self.config.is_encoder_decoder:
                        cross_attentions = _split_model_outputs(
4562
                            cross_attentions, outputs.cross_attentions, cur_len, added_len
4563
4564
4565
4566
                        )
                        decoder_attentions = _split_model_outputs(
                            decoder_attentions,
                            outputs.decoder_attentions,
4567
                            cur_len,
4568
                            added_len,
4569
4570
4571
4572
4573
4574
                            is_decoder_attention=True,
                        )
                    else:
                        decoder_attentions = _split_model_outputs(
                            decoder_attentions,
                            outputs.attentions,
4575
                            cur_len,
4576
                            added_len,
4577
4578
4579
4580
4581
                            is_decoder_attention=True,
                        )
                if output_hidden_states:
                    if self.config.is_encoder_decoder:
                        decoder_hidden_states = _split_model_outputs(
4582
                            decoder_hidden_states, outputs.decoder_hidden_states, cur_len, added_len
4583
4584
4585
                        )
                    else:
                        decoder_hidden_states = _split_model_outputs(
4586
                            decoder_hidden_states, outputs.hidden_states, cur_len, added_len
4587
4588
4589
4590
4591
4592
4593
4594
4595
                        )

            model_kwargs = self._update_model_kwargs_for_generation(
                outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
            )

            # if eos_token was found in one sentence, set sentence to finished
            if eos_token_id_tensor is not None:
                unfinished_sequences = unfinished_sequences.mul(
4596
4597
4598
4599
                    input_ids[:, -1]
                    .tile(eos_token_id_tensor.shape[0], 1)
                    .ne(eos_token_id_tensor.unsqueeze(1))
                    .prod(dim=0)
4600
4601
                )

4602
4603
                # stop when each sentence is finished
                if unfinished_sequences.max() == 0:
4604
4605
                    this_peer_finished = True

4606
4607
4608
4609
4610
4611
4612
            # stop if we exceed the maximum length
            if stopping_criteria(input_ids, scores):
                this_peer_finished = True

            if this_peer_finished and not synced_gpus:
                break

4613
4614
4615
4616
4617
4618
4619
4620
4621
4622
4623
4624
4625
4626
4627
4628
4629
4630
4631
4632
4633
4634
4635
4636
4637
4638
4639
4640
4641
4642
4643
4644
4645
4646
4647
4648
4649
4650
4651
        if streamer is not None:
            streamer.end()

        if return_dict_in_generate:
            if self.config.is_encoder_decoder:
                return GreedySearchEncoderDecoderOutput(
                    sequences=input_ids,
                    scores=scores,
                    encoder_attentions=encoder_attentions,
                    encoder_hidden_states=encoder_hidden_states,
                    decoder_attentions=decoder_attentions,
                    cross_attentions=cross_attentions,
                    decoder_hidden_states=decoder_hidden_states,
                )
            else:
                return GreedySearchDecoderOnlyOutput(
                    sequences=input_ids,
                    scores=scores,
                    attentions=decoder_attentions,
                    hidden_states=decoder_hidden_states,
                )
        else:
            return input_ids


def _crop_past_key_values(model, past_key_values, maximum_length):
    """Crops the past key values up to a certain maximum length."""
    new_past = []
    if model.config.is_encoder_decoder:
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length, :],
                    past_key_values[idx][1][:, :, :maximum_length, :],
                    past_key_values[idx][2],
                    past_key_values[idx][3],
                )
            )
        past_key_values = tuple(new_past)
4652
4653
4654
4655
    # bloom is special
    elif "bloom" in model.__class__.__name__.lower() or (
        model.config.architectures is not None and "bloom" in model.config.architectures[0].lower()
    ):
4656
4657
4658
4659
4660
4661
4662
4663
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length],
                    past_key_values[idx][1][:, :maximum_length, :],
                )
            )
        past_key_values = tuple(new_past)
4664
4665
4666
4667
    # gptbigcode is too
    elif "gptbigcode" in model.__class__.__name__.lower() or (
        model.config.architectures is not None and "gptbigcode" in model.config.architectures[0].lower()
    ):
4668
4669
4670
4671
4672
4673
        if model.config.multi_query:
            for idx in range(len(past_key_values)):
                past_key_values[idx] = past_key_values[idx][:, :maximum_length, :]
        else:
            for idx in range(len(past_key_values)):
                past_key_values[idx] = past_key_values[idx][:, :, :maximum_length, :]
4674
4675
4676
4677
4678
4679
4680
4681
4682
4683
4684
4685
    else:
        for idx in range(len(past_key_values)):
            new_past.append(
                (
                    past_key_values[idx][0][:, :, :maximum_length, :],
                    past_key_values[idx][1][:, :, :maximum_length, :],
                )
            )
        past_key_values = tuple(new_past)
    return past_key_values


4686
def _split_model_outputs(outputs, new_outputs, cur_len, added_len, is_decoder_attention=False):
4687
4688
4689
4690
4691
4692
    """
    Given the (decoder/cross attentions)/(decoder hidden states) for multiple generated tokens, splits it into a tuple
    where each member corresponds to a single generated token.
    """
    # Retrocompatibility: in our generation functions, the first iteration includes the attention/hidden states for the
    # prompt.
4693
    if len(outputs) == 0:
4694
4695
        new_tuple = ()
        for layer in new_outputs:
4696
4697
            last_dim_size = cur_len if is_decoder_attention else layer.shape[-1]
            new_tuple += (layer[..., :cur_len, :last_dim_size],)
4698
        outputs += (new_tuple,)
4699
4700
4701
        # The first iteration contains the prompt + 1 generated token, let's update the length variables accordingly
        cur_len += 1
        added_len -= cur_len
4702

4703
    for i in range(added_len):
4704
4705
        new_tuple = ()
        for layer in new_outputs:
4706
            last_dim_size = cur_len + i if is_decoder_attention else layer.shape[-1]
4707
4708
4709
4710
            new_tuple += (layer[..., i : i + 1, :last_dim_size],)
        outputs += (new_tuple,)
    return outputs

4711
4712
4713
4714
4715
4716
4717
4718
4719
4720
4721
4722
4723
4724
4725
4726
4727
4728
4729
4730
4731
4732
4733
4734
4735
4736
4737
4738
4739
4740
4741
4742
4743
4744
4745
4746
4747
4748
4749
4750
4751
4752
4753
4754
4755
4756
4757
4758
4759
4760
4761
4762
4763
4764
4765
4766
4767

def top_k_top_p_filtering(
    logits: torch.FloatTensor,
    top_k: int = 0,
    top_p: float = 1.0,
    filter_value: float = -float("Inf"),
    min_tokens_to_keep: int = 1,
) -> torch.FloatTensor:
    """
    Filter a distribution of logits using top-k and/or nucleus (top-p) filtering

    Args:
        logits: logits distribution shape (batch size, vocabulary size)
        top_k (`int`, *optional*, defaults to 0):
            If > 0, only keep the top k tokens with highest probability (top-k filtering)
        top_p (`float`, *optional*, defaults to 1.0):
            If < 1.0, only keep the top tokens with cumulative probability >= top_p (nucleus filtering). Nucleus
            filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
        min_tokens_to_keep (`int`, *optional*, defaults to 1):
            Minimumber of tokens we keep per batch example in the output.

    From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
    """
    if top_k > 0:
        logits = TopKLogitsWarper(top_k=top_k, filter_value=filter_value, min_tokens_to_keep=min_tokens_to_keep)(
            None, logits
        )

    if 0 <= top_p <= 1.0:
        logits = TopPLogitsWarper(top_p=top_p, filter_value=filter_value, min_tokens_to_keep=min_tokens_to_keep)(
            None, logits
        )

    return logits


def _ranking_fast(
    context_hidden: torch.FloatTensor,
    next_hidden: torch.FloatTensor,
    next_top_k_probs: torch.FloatTensor,
    alpha: float,
    beam_width: int,
) -> torch.FloatTensor:
    """
    Reranks the top_k candidates based on a degeneration penalty (cosine similarity with previous tokens), as described
    in the paper "A Contrastive Framework for Neural Text Generation". Returns the index of the best candidate for each
    row in the batch.
    """
    norm_context_hidden = context_hidden / context_hidden.norm(dim=2, keepdim=True)
    norm_next_hidden = next_hidden / next_hidden.norm(dim=2, keepdim=True)
    cosine_matrix = torch.matmul(norm_context_hidden, norm_next_hidden.transpose(1, 2)).squeeze(-1)  # [B*K, S]
    degeneration_penalty, _ = torch.max(cosine_matrix, dim=-1)  # [B*K]
    next_top_k_probs = next_top_k_probs.view(-1)  # [B*K]
    contrastive_score = (1.0 - alpha) * next_top_k_probs - alpha * degeneration_penalty
    contrastive_score = torch.stack(torch.split(contrastive_score, beam_width))  # [B, K]
    _, selected_idx = contrastive_score.max(dim=-1)  # [B]
    return selected_idx