test_generation_utils.py 115 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 HuggingFace Team Inc.
#
# 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 clone 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 inspect
18
19
20
import unittest

from transformers import is_torch_available
21
from transformers.testing_utils import require_torch, slow, torch_device
22

23
from ..test_modeling_common import floats_tensor, ids_tensor
24

25
26
27
28

if is_torch_available():
    import torch

29
    from transformers import (
30
31
        AutoModelForSeq2SeqLM,
        AutoTokenizer,
32
33
34
35
        BartForConditionalGeneration,
        BartTokenizer,
        GPT2LMHeadModel,
        GPT2Tokenizer,
36
        ImageGPTForCausalImageModeling,
37
38
39
        Speech2TextForConditionalGeneration,
        SpeechEncoderDecoderModel,
        VisionEncoderDecoderModel,
40
        pipeline,
41
42
        top_k_top_p_filtering,
    )
43
    from transformers.generation_beam_constraints import DisjunctiveConstraint, PhrasalConstraint
44
    from transformers.generation_beam_search import BeamSearchScorer, ConstrainedBeamSearchScorer
45
    from transformers.generation_logits_process import (
46
47
        ForcedBOSTokenLogitsProcessor,
        ForcedEOSTokenLogitsProcessor,
48
        HammingDiversityLogitsProcessor,
49
        InfNanRemoveLogitsProcessor,
50
51
52
53
54
55
56
57
58
        LogitsProcessorList,
        MinLengthLogitsProcessor,
        NoBadWordsLogitsProcessor,
        NoRepeatNGramLogitsProcessor,
        RepetitionPenaltyLogitsProcessor,
        TemperatureLogitsWarper,
        TopKLogitsWarper,
        TopPLogitsWarper,
    )
59
    from transformers.generation_stopping_criteria import MaxLengthCriteria, StoppingCriteria, StoppingCriteriaList
60
    from transformers.generation_utils import (
61
62
        BeamSampleDecoderOnlyOutput,
        BeamSampleEncoderDecoderOutput,
63
64
65
66
67
68
69
        BeamSearchDecoderOnlyOutput,
        BeamSearchEncoderDecoderOutput,
        GreedySearchDecoderOnlyOutput,
        GreedySearchEncoderDecoderOutput,
        SampleDecoderOnlyOutput,
        SampleEncoderDecoderOutput,
    )
70
71
72
73
74


class GenerationTesterMixin:
    model_tester = None
    all_generative_model_classes = ()
Suraj Patil's avatar
Suraj Patil committed
75
    input_name = "input_ids"
76
77
78

    def _get_input_ids_and_config(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
Suraj Patil's avatar
Suraj Patil committed
79
        input_ids = inputs_dict[self.input_name]
80
81
82
83
84
85
86
87
88
89
90

        # cut to half length & take max batch_size 3
        max_batch_size = 2
        sequence_length = input_ids.shape[-1] // 2
        input_ids = input_ids[:max_batch_size, :sequence_length]

        # generate max 3 tokens
        max_length = input_ids.shape[-1] + 3
        if config.eos_token_id is not None and config.pad_token_id is None:
            # hack to allow generate for models such as GPT2 as is done in `generate()`
            config.pad_token_id = config.eos_token_id
91
92
93
94
95
96
97

        # TransfoXL has no attention mask
        if "transfoxl" in config.__class__.__name__.lower():
            attention_mask = None
        else:
            attention_mask = torch.ones_like(input_ids, dtype=torch.long)[:max_batch_size, :sequence_length]

98
99
100
        return config, input_ids, attention_mask, max_length

    @staticmethod
101
102
103
104
105
106
107
108
    def _get_logits_processor_and_kwargs(
        input_length,
        eos_token_id,
        forced_bos_token_id=None,
        forced_eos_token_id=None,
        max_length=None,
        diversity_penalty=None,
    ):
109
        process_kwargs = {
110
            "min_length": input_length + 1 if max_length is None else max_length - 1,
111
112
113
114
115
116
            "bad_words_ids": [[1, 0]],
            "no_repeat_ngram_size": 2,
            "repetition_penalty": 1.2,
        }
        logits_processor = LogitsProcessorList(
            (
117
118
119
120
121
122
123
                [
                    HammingDiversityLogitsProcessor(diversity_penalty, num_beams=2, num_beam_groups=2),
                ]
                if diversity_penalty is not None
                else []
            )
            + (
124
125
126
127
128
129
                [
                    MinLengthLogitsProcessor(process_kwargs["min_length"], eos_token_id),
                ]
                if eos_token_id is not None
                else []
            )
130
131
132
133
134
135
136
137
138
139
140
141
            + (
                [
                    ForcedBOSTokenLogitsProcessor(forced_bos_token_id),
                ]
                if forced_bos_token_id is not None
                else []
            )
            + (
                [ForcedEOSTokenLogitsProcessor(max_length, forced_eos_token_id)]
                if forced_eos_token_id is not None
                else []
            )
142
143
144
145
146
147
148
149
150
151
152
153
154
            + [
                NoBadWordsLogitsProcessor(process_kwargs["bad_words_ids"], eos_token_id),
                NoRepeatNGramLogitsProcessor(process_kwargs["no_repeat_ngram_size"]),
                RepetitionPenaltyLogitsProcessor(process_kwargs["repetition_penalty"]),
            ]
        )
        return process_kwargs, logits_processor

    @staticmethod
    def _get_warper_and_kwargs(num_beams):
        warp_kwargs = {"top_k": 10, "top_p": 0.7, "temperature": 0.7}
        logits_warper = LogitsProcessorList(
            [
Patrick von Platen's avatar
Patrick von Platen committed
155
                TemperatureLogitsWarper(warp_kwargs["temperature"]),
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
                TopKLogitsWarper(top_k=warp_kwargs["top_k"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
                TopPLogitsWarper(top_p=warp_kwargs["top_p"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
            ]
        )
        return warp_kwargs, logits_warper

    @staticmethod
    def _get_beam_scorer_and_kwargs(batch_size, max_length, num_return_sequences=1):
        beam_kwargs = {
            "early_stopping": False,
            "length_penalty": 2.0,
            "num_beams": 2,
            "num_return_sequences": num_return_sequences,
        }
        beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=beam_kwargs["num_beams"],
            device=torch_device,
            length_penalty=beam_kwargs["length_penalty"],
            do_early_stopping=beam_kwargs["early_stopping"],
            num_beam_hyps_to_keep=num_return_sequences,
        )
        return beam_kwargs, beam_scorer

180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
    @staticmethod
    def _get_diverse_beam_scorer_and_kwargs(batch_size, max_length, num_return_sequences=1):
        beam_kwargs = {
            "early_stopping": False,
            "length_penalty": 2.0,
            "num_beams": 2,
            "num_return_sequences": num_return_sequences,
            "num_beam_groups": 2,  # one beam per group
            "diversity_penalty": 2.0,
        }
        beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=beam_kwargs["num_beams"],
            device=torch_device,
            length_penalty=beam_kwargs["length_penalty"],
            do_early_stopping=beam_kwargs["early_stopping"],
            num_beam_hyps_to_keep=num_return_sequences,
            num_beam_groups=beam_kwargs["num_beam_groups"],
        )
        return beam_kwargs, beam_scorer

201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
    @staticmethod
    def _get_constrained_beam_scorer_and_kwargs(batch_size, max_length, constraints, num_return_sequences=1):
        beam_kwargs = {
            "early_stopping": False,
            "length_penalty": 2.0,
            "num_beams": num_return_sequences * 4,
            "num_return_sequences": num_return_sequences,
        }
        beam_scorer = ConstrainedBeamSearchScorer(
            batch_size=batch_size,
            constraints=constraints,
            num_beams=beam_kwargs["num_beams"],
            device=torch_device,
            length_penalty=beam_kwargs["length_penalty"],
            do_early_stopping=beam_kwargs["early_stopping"],
            num_beam_hyps_to_keep=num_return_sequences,
        )
        return beam_kwargs, beam_scorer

220
    @staticmethod
221
222
223
    def _get_encoder_outputs(
        model, input_ids, attention_mask, output_attentions=None, output_hidden_states=None, num_interleave=1
    ):
224
        encoder = model.get_encoder()
225
226
227
228
229
230
        encoder_outputs = encoder(
            input_ids,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
        )
231
232
233
234
235
236
237
        encoder_outputs["last_hidden_state"] = encoder_outputs.last_hidden_state.repeat_interleave(
            num_interleave, dim=0
        )
        input_ids = torch.zeros_like(input_ids[:, :1]) + model._get_decoder_start_token_id()
        attention_mask = None
        return encoder_outputs, input_ids, attention_mask

238
239
240
241
242
243
244
245
246
247
248
    def _greedy_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
249
250
        if model.config.is_encoder_decoder:
            max_length = 4
251
        logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
252
253
254
255
256
            input_ids.shape[-1],
            eos_token_id=model.config.eos_token_id,
            forced_bos_token_id=model.config.forced_bos_token_id,
            forced_eos_token_id=model.config.forced_eos_token_id,
            max_length=max_length,
257
258
259
        )

        kwargs = {}
260
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
261
262
263
264
265
266
267
268
269
        output_generate = model.generate(
            input_ids,
            do_sample=False,
            num_beams=1,
            max_length=max_length,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_scores=output_scores,
            return_dict_in_generate=return_dict_in_generate,
270
            remove_invalid_values=True,
271
            **logits_process_kwargs,
272
            **model_kwargs,
273
274
275
276
277
278
279
280
281
282
283
284
285
        )

        if model.config.is_encoder_decoder:
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
                model,
                input_ids,
                attention_mask,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs

        with torch.no_grad():
286
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
287
288
289
290
291
292
293
294
295
            output_greedy = model.greedy_search(
                input_ids,
                max_length=max_length,
                logits_processor=logits_processor,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                output_scores=output_scores,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
296
                **model_kwargs,
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
            )
        return output_greedy, output_generate

    def _sample_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        num_return_sequences,
        logits_processor,
        logits_warper,
        logits_warper_kwargs,
        process_kwargs,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
        torch.manual_seed(0)
317
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
318
319
320
321
322
323
324
325
326
327
        output_generate = model.generate(
            input_ids,
            do_sample=True,
            num_beams=1,
            max_length=max_length,
            num_return_sequences=num_return_sequences,
            output_scores=output_scores,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict_in_generate=return_dict_in_generate,
328
            remove_invalid_values=True,
329
330
            **logits_warper_kwargs,
            **process_kwargs,
331
            **model_kwargs,
332
333
334
335
336
        )

        torch.manual_seed(0)
        kwargs = {}
        if model.config.is_encoder_decoder:
337
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
338
339
340
341
342
343
344
345
                model,
                input_ids,
                attention_mask,
                num_interleave=num_return_sequences,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs
346
347
        elif attention_mask is not None:
            attention_mask = attention_mask.repeat_interleave(num_return_sequences, dim=0)
348

349
350
351
        # prevent flaky generation test failures
        logits_processor.append(InfNanRemoveLogitsProcessor())

352
        with torch.no_grad():
353
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
Vasudev Gupta's avatar
Vasudev Gupta committed
354
            output_sample = model.sample(
355
                input_ids.repeat_interleave(num_return_sequences, dim=0),
Vasudev Gupta's avatar
Vasudev Gupta committed
356
357
358
359
360
361
362
363
                max_length=max_length,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                output_scores=output_scores,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
364
                **model_kwargs,
Vasudev Gupta's avatar
Vasudev Gupta committed
365
            )
366

367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
        return output_sample, output_generate

    def _beam_search_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        beam_scorer,
        beam_kwargs,
        logits_processor,
        logits_process_kwargs,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
384
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
385
386
387
388
389
390
391
392
        output_generate = model.generate(
            input_ids,
            do_sample=False,
            max_length=max_length,
            output_scores=output_scores,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict_in_generate=return_dict_in_generate,
393
            remove_invalid_values=True,
394
395
            **beam_kwargs,
            **logits_process_kwargs,
396
            **model_kwargs,
397
398
399
400
401
        )

        # beam_search does not automatically interleave `batch_size` dim for `num_beams`
        kwargs = {}
        if model.config.is_encoder_decoder:
402
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
403
404
405
406
407
408
409
410
                model,
                input_ids,
                attention_mask,
                num_interleave=beam_scorer.num_beams,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs
411
412
        elif attention_mask is not None:
            attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams, dim=0)
413
414

        with torch.no_grad():
415
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
416
            output_beam_search = model.beam_search(
417
                input_ids.repeat_interleave(beam_scorer.num_beams, dim=0),
418
419
420
421
422
423
424
425
                beam_scorer,
                max_length=max_length,
                logits_processor=logits_processor,
                output_scores=output_scores,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
426
                **model_kwargs,
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
            )
        return output_generate, output_beam_search

    def _beam_sample_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        num_return_sequences,
        beam_scorer,
        beam_kwargs,
        logits_warper,
        logits_warper_kwargs,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
        torch.manual_seed(0)
447
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
448
449
450
451
452
453
454
455
        output_generate = model.generate(
            input_ids,
            do_sample=True,
            max_length=max_length,
            output_scores=output_scores,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict_in_generate=return_dict_in_generate,
456
            remove_invalid_values=True,
457
458
            **beam_kwargs,
            **logits_warper_kwargs,
459
            **model_kwargs,
460
461
462
463
464
465
466
467
468
469
470
471
472
        )
        # beam_search does not automatically interleave `batch_size` dim for `num_beams * num_return_sequences`
        kwargs = {}
        if model.config.is_encoder_decoder:
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
                model,
                input_ids,
                attention_mask,
                num_interleave=beam_scorer.num_beams * num_return_sequences,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs
473
        elif attention_mask is not None:
474
475
            attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams * num_return_sequences, dim=0)

476
477
478
479
        # prevent flaky generation test failures
        logits_processor = LogitsProcessorList()
        logits_processor.append(InfNanRemoveLogitsProcessor())

480
481
        torch.manual_seed(0)
        with torch.no_grad():
482
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
483
484
485
486
487
            output_beam_sample = model.beam_sample(
                input_ids.repeat_interleave(beam_scorer.num_beams * num_return_sequences, dim=0),
                beam_scorer,
                max_length=max_length,
                logits_warper=logits_warper,
488
                logits_processor=logits_processor,
489
490
491
492
493
                output_scores=output_scores,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
494
                **model_kwargs,
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
            )

        return output_generate, output_beam_sample

    def _group_beam_search_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        beam_scorer,
        beam_kwargs,
        logits_processor,
        logits_process_kwargs,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
514
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
515
516
517
518
519
520
521
522
        output_generate = model.generate(
            input_ids,
            do_sample=False,
            max_length=max_length,
            output_scores=output_scores,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict_in_generate=return_dict_in_generate,
523
            remove_invalid_values=True,
524
525
            **beam_kwargs,
            **logits_process_kwargs,
526
            **model_kwargs,
527
528
529
530
531
        )

        # group_beam_search does not automatically interleave `batch_size` dim for `num_beams`
        kwargs = {}
        if model.config.is_encoder_decoder:
532
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
533
534
535
536
537
538
539
540
                model,
                input_ids,
                attention_mask,
                num_interleave=beam_scorer.num_beams,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs
541
542
        elif attention_mask is not None:
            attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams, dim=0)
543
544

        with torch.no_grad():
545
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
546
            output_group_beam_search = model.group_beam_search(
547
                input_ids.repeat_interleave(beam_scorer.num_beams, dim=0),
548
549
550
551
552
553
554
555
                beam_scorer,
                max_length=max_length,
                logits_processor=logits_processor,
                output_scores=output_scores,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
556
                **model_kwargs,
557
558
559
            )
        return output_generate, output_group_beam_search

560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
    def _constrained_beam_search_generate(
        self,
        model,
        input_ids,
        attention_mask,
        max_length,
        constrained_beam_scorer,
        constraints,
        beam_kwargs,
        logits_processor,
        logits_process_kwargs,
        output_scores=False,
        output_attentions=False,
        output_hidden_states=False,
        return_dict_in_generate=False,
    ):
576
        model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
577
578
579
580
581
582
583
584
585
586
587
588
        output_generate = model.generate(
            input_ids,
            do_sample=False,
            max_length=max_length,
            output_scores=output_scores,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict_in_generate=return_dict_in_generate,
            remove_invalid_values=True,
            constraints=constraints,
            **beam_kwargs,
            **logits_process_kwargs,
589
            **model_kwargs,
590
591
592
593
594
        )

        # group_beam_search does not automatically interleave `batch_size` dim for `num_beams`
        kwargs = {}
        if model.config.is_encoder_decoder:
595
            encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
596
597
598
599
600
601
602
603
                model,
                input_ids,
                attention_mask,
                num_interleave=constrained_beam_scorer.num_beams,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
            )
            kwargs["encoder_outputs"] = encoder_outputs
604
605
        elif attention_mask is not None:
            attention_mask = attention_mask.repeat_interleave(constrained_beam_scorer.num_beams, dim=0)
606
607

        with torch.no_grad():
608
            model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
609
            output_group_beam_search = model.constrained_beam_search(
610
                input_ids.repeat_interleave(constrained_beam_scorer.num_beams, dim=0),
611
612
613
614
615
616
617
618
                constrained_beam_scorer,
                max_length=max_length,
                logits_processor=logits_processor,
                output_scores=output_scores,
                output_attentions=output_attentions,
                output_hidden_states=output_hidden_states,
                return_dict_in_generate=return_dict_in_generate,
                **kwargs,
619
                **model_kwargs,
620
621
622
            )
        return output_generate, output_group_beam_search

623
    def test_greedy_generate(self):
624
        # check `generate()` and `greedy_search()` are equal
625
626
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
627
628
629
630
            # test old generation output for backwards compatibility
            model = model_class(config).to(torch_device).eval()
            output_greedy, output_generate = self._greedy_generate(
                model=model, input_ids=input_ids, attention_mask=attention_mask, max_length=max_length
631
            )
632
            self.assertListEqual(output_greedy.tolist(), output_generate.tolist())
633

634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
    def test_greedy_generate_dict_outputs(self):
        for model_class in self.all_generative_model_classes:
            # disable cache
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
            config.use_cache = False
            model = model_class(config).to(torch_device).eval()
            output_greedy, output_generate = self._greedy_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
            )
650
651

            if model.config.is_encoder_decoder:
652
653
654
655
656
                self.assertIsInstance(output_greedy, GreedySearchEncoderDecoderOutput)
                self.assertIsInstance(output_generate, GreedySearchEncoderDecoderOutput)
            else:
                self.assertIsInstance(output_greedy, GreedySearchDecoderOnlyOutput)
                self.assertIsInstance(output_generate, GreedySearchDecoderOnlyOutput)
657

658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
            self.assertListEqual(output_generate.sequences.tolist(), output_greedy.sequences.tolist())

            for output in (output_greedy, output_generate):
                self._check_outputs(output, input_ids, model.config)

    def test_greedy_generate_dict_outputs_use_cache(self):
        for model_class in self.all_generative_model_classes:
            # enable cache
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()

            if not hasattr(config, "use_cache"):
                # only relevant if model has "use_cache"
                return

            config.use_cache = True
673
            config.is_decoder = True
674
675
676
677
            model = model_class(config).to(torch_device).eval()
            output_greedy, output_generate = self._greedy_generate(
                model=model,
                input_ids=input_ids,
678
679
                attention_mask=attention_mask,
                max_length=max_length,
680
681
682
683
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
684
            )
685

686
687
688
689
            self.assertListEqual(output_generate.sequences.tolist(), output_greedy.sequences.tolist())

            for output in (output_greedy, output_generate):
                self._check_outputs(output, input_ids, model.config, use_cache=True)
690
691
692
693

    def test_sample_generate(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
694
            model = model_class(config).to(torch_device).eval()
695
696
697
698

            if model.config.is_encoder_decoder:
                max_length = 4

699
700
701
702
703
704
705
706
707
            process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
                input_ids.shape[-1],
                model.config.eos_token_id,
                forced_bos_token_id=model.config.forced_bos_token_id,
                forced_eos_token_id=model.config.forced_eos_token_id,
                max_length=max_length,
            )
            logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)

708
709
710
711
712
            # check `generate()` and `sample()` are equal
            output_sample, output_generate = self._sample_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
713
                max_length=max_length,
714
715
716
717
718
719
720
721
722
723
724
725
                num_return_sequences=1,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                logits_warper_kwargs=logits_warper_kwargs,
                process_kwargs=process_kwargs,
            )
            self.assertListEqual(output_sample.tolist(), output_generate.tolist())

            # check `generate()` and `sample()` yield equal results for `num_return_sequences`
            output_sample, output_generate = self._sample_generate(
                model=model,
                input_ids=input_ids,
726
                attention_mask=attention_mask,
727
728
729
730
731
732
                max_length=max_length,
                num_return_sequences=3,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                logits_warper_kwargs=logits_warper_kwargs,
                process_kwargs=process_kwargs,
733
            )
734
            self.assertListEqual(output_sample.tolist(), output_generate.tolist())
735

736
737
738
739
740
741
    def test_sample_generate_dict_output(self):
        for model_class in self.all_generative_model_classes:
            # disable cache
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
            config.use_cache = False
            model = model_class(config).to(torch_device).eval()
742
743
744
            if model.config.is_encoder_decoder:
                max_length = 4

745
            process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
746
747
748
749
750
                input_ids.shape[-1],
                model.config.eos_token_id,
                forced_bos_token_id=model.config.forced_bos_token_id,
                forced_eos_token_id=model.config.forced_eos_token_id,
                max_length=max_length,
751
752
            )
            logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)
753

754
755
756
            output_sample, output_generate = self._sample_generate(
                model=model,
                input_ids=input_ids,
757
                attention_mask=attention_mask,
758
759
760
761
762
763
764
765
766
767
                max_length=max_length,
                num_return_sequences=2,
                logits_processor=logits_processor,
                logits_warper=logits_warper,
                logits_warper_kwargs=logits_warper_kwargs,
                process_kwargs=process_kwargs,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
768
769
770
            )

            if model.config.is_encoder_decoder:
771
772
                self.assertIsInstance(output_sample, SampleEncoderDecoderOutput)
                self.assertIsInstance(output_generate, SampleEncoderDecoderOutput)
773
            else:
774
775
776
777
778
779
780
                self.assertIsInstance(output_sample, SampleDecoderOnlyOutput)
                self.assertIsInstance(output_generate, SampleDecoderOnlyOutput)

            self.assertListEqual(output_generate.sequences.tolist(), output_sample.sequences.tolist())

            for output in (output_sample, output_generate):
                self._check_outputs(output, input_ids, model.config, num_return_sequences=2)
781
782
783
784

    def test_beam_search_generate(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
785
786
787
788
789

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
790
            config.forced_eos_token_id = None
791

792
            model = model_class(config).to(torch_device).eval()
793
794
            if model.config.is_encoder_decoder:
                max_length = 4
795
796

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
797
798
799
800
801
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
802
803
            )
            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
804
805
806
807
808

            # check `generate()` and `beam_search()` are equal
            output_generate, output_beam_search = self._beam_search_generate(
                model=model,
                input_ids=input_ids,
809
810
                attention_mask=attention_mask,
                max_length=max_length,
811
812
813
814
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_process_kwargs=logits_process_kwargs,
                logits_processor=logits_processor,
815
            )
816

817
            self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
818
819
820
821
822
823
824
825
826

            # check `generate()` and `beam_search()` are equal for `num_return_sequences`
            num_return_sequences = 2
            if model.config.is_encoder_decoder:
                max_length = 4
            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, num_return_sequences=num_return_sequences
            )

827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
            output_generate, output_beam_search = self._beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_process_kwargs=logits_process_kwargs,
                logits_processor=logits_processor,
            )
            self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())

    def test_beam_search_generate_dict_output(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
842
843

            # disable cache
844
            config.use_cache = False
845
846
847
848
849

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
850
            config.forced_eos_token_id = None
851

852
853
854
            model = model_class(config).to(torch_device).eval()
            if model.config.is_encoder_decoder:
                max_length = 4
855
856
857
858
859
860
861
862

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
            )
863
864
865
866
            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
            output_generate, output_beam_search = self._beam_search_generate(
                model=model,
                input_ids=input_ids,
867
868
                attention_mask=attention_mask,
                max_length=max_length,
869
870
871
872
873
874
875
876
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_process_kwargs=logits_process_kwargs,
                logits_processor=logits_processor,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
877
878
            )
            if model.config.is_encoder_decoder:
879
880
                self.assertIsInstance(output_beam_search, BeamSearchEncoderDecoderOutput)
                self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
881
            else:
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
                self.assertIsInstance(output_beam_search, BeamSearchDecoderOnlyOutput)
                self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)

            self.assertListEqual(output_generate.sequences.tolist(), output_beam_search.sequences.tolist())
            self.assertTrue(
                torch.allclose(output_generate["sequences_scores"], output_beam_search["sequences_scores"], atol=1e-3)
            )
            self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
            self.assertTrue((output_generate["sequences_scores"] < 0).all().item())

            for output in (output_beam_search, output_generate):
                self._check_outputs(output, input_ids, model.config, num_return_sequences=beam_scorer.num_beams)

    def test_beam_search_generate_dict_outputs_use_cache(self):
        for model_class in self.all_generative_model_classes:
            # enable cache
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()

900
901
902
903
            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
904
            config.forced_eos_token_id = None
905

906
907
908
909
910
            if not hasattr(config, "use_cache"):
                # only relevant if model has "use_cache"
                return

            model = model_class(config).to(torch_device).eval()
911
912
            if model.config.is_encoder_decoder:
                max_length = 4
913
914

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
915
916
917
918
919
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
920
921
922
923
924
            )

            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)

            config.use_cache = True
925
            config.is_decoder = True
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
            model = model_class(config).to(torch_device).eval()
            output_beam, output_generate = self._beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_process_kwargs=logits_process_kwargs,
                logits_processor=logits_processor,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
            )

            self.assertListEqual(output_generate.sequences.tolist(), output_beam.sequences.tolist())

            for output in (output_beam, output_generate):
                self._check_outputs(
                    output, input_ids, model.config, use_cache=True, num_return_sequences=beam_scorer.num_beams
947
948
949
950
951
                )

    def test_beam_sample_generate(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
952
953
954
955
956

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
957
            config.forced_eos_token_id = None
958

959
960
            logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)

961
            model = model_class(config).to(torch_device).eval()
962
963
964
965
966
967
968
969
970
971

            # check `generate()` and `beam_search()` are equal
            # change `num_return_sequences = 2` but not for `beam_scorer`
            num_return_sequences = 2
            if model.config.is_encoder_decoder:
                max_length = 4
            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(
                input_ids.shape[0] * num_return_sequences, max_length
            )
            beam_kwargs["num_return_sequences"] = num_return_sequences
972
973
974
975

            output_generate, output_beam_sample = self._beam_sample_generate(
                model=model,
                input_ids=input_ids,
976
977
                attention_mask=attention_mask,
                max_length=max_length,
978
979
980
981
982
                num_return_sequences=num_return_sequences,
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_warper=logits_warper,
                logits_warper_kwargs=logits_warper_kwargs,
983
            )
984
985
986
987
988
            self.assertListEqual(output_generate.tolist(), output_beam_sample.tolist())

    def test_beam_sample_generate_dict_output(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
989
990

            # disable cache
991
            config.use_cache = False
992
993
994
995
996

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
997
            config.forced_eos_token_id = None
998

999
1000
1001
1002
            model = model_class(config).to(torch_device).eval()
            logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)

            num_return_sequences = 2
1003
            if model.config.is_encoder_decoder:
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
                max_length = 4
            beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(
                input_ids.shape[0] * num_return_sequences, max_length
            )
            beam_kwargs["num_return_sequences"] = num_return_sequences

            output_beam_sample, output_generate = self._beam_sample_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                num_return_sequences=num_return_sequences,
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_warper=logits_warper,
                logits_warper_kwargs=logits_warper_kwargs,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
            )

            if model.config.is_encoder_decoder:
1027
1028
                self.assertIsInstance(output_beam_sample, BeamSampleEncoderDecoderOutput)
                self.assertIsInstance(output_generate, BeamSampleEncoderDecoderOutput)
1029
            else:
1030
1031
                self.assertIsInstance(output_beam_sample, BeamSampleDecoderOnlyOutput)
                self.assertIsInstance(output_generate, BeamSampleDecoderOnlyOutput)
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042

            self.assertListEqual(output_generate.sequences.tolist(), output_beam_sample.sequences.tolist())
            self.assertTrue(
                torch.allclose(output_generate["sequences_scores"], output_beam_sample["sequences_scores"], atol=1e-3)
            )
            self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
            self.assertTrue((output_generate["sequences_scores"] < 0).all().item())

            for output in (output_beam_sample, output_generate):
                self._check_outputs(
                    output, input_ids, model.config, num_return_sequences=num_return_sequences * beam_scorer.num_beams
1043
1044
                )

1045
1046
    def test_generate_without_input_ids(self):
        config, _, _, max_length = self._get_input_ids_and_config()
1047

1048
1049
1050
        # if no bos token id => cannot generate from None
        if config.bos_token_id is None:
            return
1051

1052
1053
1054
        for model_class in self.all_generative_model_classes:
            model = model_class(config).to(torch_device)
            model.eval()
1055

1056
            output_ids_generate = model.generate(do_sample=False, max_length=max_length, remove_invalid_values=True)
1057
            self.assertIsNotNone(output_ids_generate)
1058

1059
1060
1061
1062
    def test_group_beam_search_generate(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()

1063
1064
1065
1066
            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
1067
1068
1069
1070
1071
            config.forced_eos_token_id = None

            model = model_class(config).to(torch_device).eval()
            if model.config.is_encoder_decoder:
                max_length = 4
1072

1073
            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
1074
1075
1076
1077
1078
1079
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
                diversity_penalty=2.0,
1080
1081
1082
1083
            )

            # check `generate()` and `group_beam_search()` are equal
            beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
1084
1085
1086
            output_generate, output_group_beam_search = self._group_beam_search_generate(
                model=model,
                input_ids=input_ids,
1087
1088
                attention_mask=attention_mask,
                max_length=max_length,
1089
1090
1091
1092
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
1093
            )
1094
            self.assertListEqual(output_generate.tolist(), output_group_beam_search.tolist())
1095
1096
1097
1098
1099
1100
1101
1102

            # check `generate()` and `group_beam_search()` are equal for `num_return_sequences`
            num_return_sequences = 2
            if model.config.is_encoder_decoder:
                max_length = 4
            beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, num_return_sequences=num_return_sequences
            )
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
            output_generate, output_group_beam_search = self._group_beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
            )
            self.assertListEqual(output_generate.tolist(), output_group_beam_search.tolist())
1114

1115
1116
1117
1118
    def test_group_beam_search_generate_dict_output(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
            config.use_cache = False
1119
1120
1121
1122
1123

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
1124
            config.forced_eos_token_id = None
1125

1126
            model = model_class(config).to(torch_device).eval()
1127
1128
            if model.config.is_encoder_decoder:
                max_length = 4
1129
1130

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
1131
1132
1133
1134
1135
1136
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
                diversity_penalty=2.0,
1137
1138
1139
1140
1141
1142
1143
1144
1145
            )

            num_return_sequences = 1
            beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, num_return_sequences=num_return_sequences
            )
            output_generate, output_group_beam_search = self._group_beam_search_generate(
                model=model,
                input_ids=input_ids,
1146
1147
                attention_mask=attention_mask,
                max_length=max_length,
1148
1149
1150
1151
1152
1153
1154
1155
                beam_scorer=beam_scorer,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
1156
1157
            )
            if model.config.is_encoder_decoder:
1158
1159
                self.assertIsInstance(output_group_beam_search, BeamSearchEncoderDecoderOutput)
                self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
1160
            else:
1161
1162
1163
1164
1165
1166
1167
                self.assertIsInstance(output_group_beam_search, BeamSearchDecoderOnlyOutput)
                self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)

            self.assertListEqual(output_generate.sequences.tolist(), output_group_beam_search.sequences.tolist())
            self.assertTrue(
                torch.allclose(
                    output_generate["sequences_scores"], output_group_beam_search["sequences_scores"], atol=1e-3
1168
                )
1169
1170
1171
1172
1173
1174
1175
1176
1177
            )
            self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
            self.assertTrue((output_generate["sequences_scores"] < 0).all().item())

            for output in (output_group_beam_search, output_generate):
                self._check_outputs(
                    output, input_ids, model.config, num_return_sequences=num_return_sequences * beam_scorer.num_beams
                )

1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
    def test_constrained_beam_search_generate(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
            config.forced_eos_token_id = None

            model = model_class(config).to(torch_device).eval()
            max_length = 20

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
            )

            # check `generate()` and `constrained_beam_search()` are equal
            # Sample constraints
            if not input_ids.dtype == torch.float32:
                min_id = torch.min(input_ids) + 3
                max_id = torch.max(input_ids)
            else:
                # otherwise this throws an error for Speech2TextModel since its inputs are floating points
                min_id = 3
                max_id = 100

1209
            force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
            constraints = [
                PhrasalConstraint(force_tokens),
            ]

            beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, constraints, num_return_sequences=1
            )
            output_generate, output_beam_search = self._constrained_beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                constrained_beam_scorer=beam_scorer,
                constraints=constraints,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
            )
            self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
            for generation_output in output_generate:
                self._check_sequence_inside_sequence(force_tokens, generation_output)

            # check `generate()` and `constrained_beam_search()` are equal for `num_return_sequences`
            # Sample constraints
1234
            force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
            constraints = [
                PhrasalConstraint(force_tokens),
            ]

            num_return_sequences = 2
            max_length = 20

            beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, constraints, num_return_sequences=num_return_sequences
            )

            output_generate, output_beam_search = self._constrained_beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                constrained_beam_scorer=beam_scorer,
                constraints=constraints,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
            )
            self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())

            for generation_output in output_generate:
                self._check_sequence_inside_sequence(force_tokens, generation_output)

    def test_constrained_beam_search_generate_dict_output(self):
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()

            # disable cache
            config.use_cache = False

            # It is important set set the eos_token_id to None to ensure that no sequences
            # shorter than `max_length` can be generated which could lead to flaky circle ci
            # failures if the top `num_return_sequences` beams are all shorter than the longest beam
            config.eos_token_id = None
            config.forced_eos_token_id = None

            model = model_class(config).to(torch_device).eval()
            if model.config.is_encoder_decoder:
                max_length = 20

            logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
                input_ids.shape[-1],
                config.eos_token_id,
                config.forced_bos_token_id,
                config.forced_eos_token_id,
                max_length,
            )

            # Sample constraints
            if not input_ids.dtype == torch.float32:
                min_id = torch.min(input_ids) + 3
                max_id = torch.max(input_ids)
            else:
                # otherwise this throws an error for Speech2TextModel since its inputs are floating points
                min_id = 3
                max_id = 100
1295
            force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
            constraints = [
                PhrasalConstraint(force_tokens),
            ]

            beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
                input_ids.shape[0], max_length, constraints, num_return_sequences=1
            )
            output_generate, output_beam_search = self._constrained_beam_search_generate(
                model=model,
                input_ids=input_ids,
                attention_mask=attention_mask,
                max_length=max_length,
                constrained_beam_scorer=beam_scorer,
                constraints=constraints,
                beam_kwargs=beam_kwargs,
                logits_processor=logits_processor,
                logits_process_kwargs=logits_process_kwargs,
                output_scores=True,
                output_hidden_states=True,
                output_attentions=True,
                return_dict_in_generate=True,
            )

            if model.config.is_encoder_decoder:
                self.assertIsInstance(output_beam_search, BeamSearchEncoderDecoderOutput)
                self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
            else:
                self.assertIsInstance(output_beam_search, BeamSearchDecoderOnlyOutput)
                self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)

            self.assertListEqual(output_generate.sequences.tolist(), output_beam_search.sequences.tolist())
            self.assertTrue(
                torch.allclose(output_generate["sequences_scores"], output_beam_search["sequences_scores"], atol=1e-3)
            )
            self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
            self.assertTrue((output_generate["sequences_scores"] < 0).all().item())

            for output in (output_beam_search, output_generate):
                self._check_outputs(output, input_ids, model.config, num_return_sequences=beam_scorer.num_beams)

1336
1337
1338
1339
1340
    def test_generate_with_head_masking(self):
        """Test designed for encoder-decoder models to ensure the attention head masking is used."""
        attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"]
        for model_class in self.all_generative_model_classes:
            config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
1341
            model = model_class(config).to(torch_device)
1342
1343
1344
1345
1346
            # We want to test only encoder-decoder models
            if not config.is_encoder_decoder:
                continue

            head_masking = {
1347
1348
1349
1350
1351
1352
1353
                "head_mask": torch.zeros(config.encoder_layers, config.encoder_attention_heads, device=torch_device),
                "decoder_head_mask": torch.zeros(
                    config.decoder_layers, config.decoder_attention_heads, device=torch_device
                ),
                "cross_attn_head_mask": torch.zeros(
                    config.decoder_layers, config.decoder_attention_heads, device=torch_device
                ),
1354
1355
1356
1357
            }

            signature = inspect.signature(model.forward)
            # We want to test only models where encoder/decoder head masking is implemented
1358
            if not set(head_masking.keys()) < set([*signature.parameters.keys()]):
1359
1360
1361
1362
1363
                continue

            for attn_name, (name, mask) in zip(attention_names, head_masking.items()):
                out = model.generate(
                    input_ids,
1364
                    attention_mask=attention_mask,
1365
1366
1367
                    num_beams=1,
                    output_attentions=True,
                    return_dict_in_generate=True,
1368
                    remove_invalid_values=True,
1369
1370
1371
1372
1373
1374
                    **{name: mask},
                )
                # We check the state of decoder_attentions and cross_attentions just from the last step
                attn_weights = out[attn_name] if attn_name == attention_names[0] else out[attn_name][-1]
                self.assertEqual(sum([w.sum().item() for w in attn_weights]), 0.0)

1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
    def _check_outputs(self, output, input_ids, config, use_cache=False, num_return_sequences=1):
        batch_size, seq_length = input_ids.shape
        num_sequences_in_output = batch_size * num_return_sequences
        gen_len = (
            output.sequences.shape[-1] - 1 if config.is_encoder_decoder else output.sequences.shape[-1] - seq_length
        )

        # scores
        self._check_scores(num_sequences_in_output, output.scores, length=gen_len, config=config)

        # Attentions
        if config.is_encoder_decoder:
            # encoder
1388
            self._check_encoder_attention_for_generate(output.encoder_attentions, batch_size, config, seq_length)
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
            # decoder
            self._check_attentions_for_generate(
                num_sequences_in_output,
                output.decoder_attentions,
                min_length=1,
                max_length=output.sequences.shape[-1],
                config=config,
                use_cache=use_cache,
            )
        else:
            # if use_cache first input is equal to no use_cache, so skip here
            attentions = output.attentions if not use_cache else output.attentions[1:]
            min_length = seq_length if not use_cache else seq_length + 1
            self._check_attentions_for_generate(
                num_sequences_in_output,
                attentions=attentions,
                min_length=min_length,
                max_length=output.sequences.shape[-1],
                config=config,
                use_cache=use_cache,
            )

        # Hidden States
        if config.is_encoder_decoder:
            # encoder
1414
1415
            self._check_encoder_hidden_states_for_generate(
                output.encoder_hidden_states, batch_size, config, seq_length
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
            )

            # decoder
            self._check_hidden_states_for_generate(
                num_sequences_in_output,
                output.decoder_hidden_states,
                min_length=1,
                max_length=output.sequences.shape[-1],
                config=config,
                use_cache=use_cache,
            )
        else:
            # if use_cache first input is equal to no use_cache, so skip here
            hidden_states = output.hidden_states if not use_cache else output.hidden_states[1:]
            min_length = seq_length if not use_cache else seq_length + 1
            self._check_hidden_states_for_generate(
                num_sequences_in_output,
                hidden_states,
                min_length=min_length,
                max_length=output.sequences.shape[-1],
                config=config,
                use_cache=use_cache,
            )

    def _check_scores(self, batch_size, scores, length, config):
        expected_shape = (batch_size, config.vocab_size)
        self.assertIsInstance(scores, tuple)
        self.assertEqual(len(scores), length)
        self.assertListEqual([iter_scores.shape for iter_scores in scores], [expected_shape] * len(scores))

    def _check_attentions_for_generate(
        self, batch_size, attentions, min_length, max_length, config, use_cache=False, num_beam_groups=1
    ):
        self.assertIsInstance(attentions, tuple)
        self.assertListEqual(
            [isinstance(iter_attentions, tuple) for iter_attentions in attentions], [True] * len(attentions)
        )
        self.assertEqual(len(attentions), (max_length - min_length) * num_beam_groups)

        for idx, iter_attentions in enumerate(attentions):
            tgt_len = min_length + idx if not use_cache else 1
            src_len = min_length + idx

            expected_shape = (
                batch_size * num_beam_groups,
                config.num_attention_heads,
                tgt_len,
                src_len,
            )
            # check attn size
            self.assertListEqual(
                [layer_attention.shape for layer_attention in iter_attentions], [expected_shape] * len(iter_attentions)
            )

1470
1471
1472
1473
1474
1475
1476
1477
    def _check_encoder_attention_for_generate(self, attentions, batch_size, config, seq_length):
        encoder_expected_shape = (batch_size, config.num_attention_heads, seq_length, seq_length)
        self.assertIsInstance(attentions, tuple)
        self.assertListEqual(
            [layer_attentions.shape for layer_attentions in attentions],
            [encoder_expected_shape] * len(attentions),
        )

1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
    def _check_hidden_states_for_generate(
        self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
    ):
        self.assertIsInstance(hidden_states, tuple)
        self.assertListEqual(
            [isinstance(iter_hidden_states, tuple) for iter_hidden_states in hidden_states],
            [True] * len(hidden_states),
        )
        self.assertEqual(len(hidden_states), (max_length - min_length) * num_beam_groups)

        for idx, iter_hidden_states in enumerate(hidden_states):
            seq_len = min_length + idx if not use_cache else 1
            expected_shape = (batch_size * num_beam_groups, seq_len, config.hidden_size)
            # check hidden size
            self.assertListEqual(
                [layer_hidden_states.shape for layer_hidden_states in iter_hidden_states],
                [expected_shape] * len(iter_hidden_states),
            )
1496

1497
1498
1499
1500
1501
1502
1503
1504
    def _check_encoder_hidden_states_for_generate(self, hidden_states, batch_size, config, seq_length):
        encoder_expected_shape = (batch_size, seq_length, config.hidden_size)
        self.assertIsInstance(hidden_states, tuple)
        self.assertListEqual(
            [layer_hidden_states.shape for layer_hidden_states in hidden_states],
            [encoder_expected_shape] * len(hidden_states),
        )

1505
    def _check_sequence_inside_sequence(self, tensor_1, tensor_2):
1506
        # check if tensor_1 inside tensor_2 or tensor_2 inside tensor_1.
1507
1508
        # set to same device. we don't care what device.

1509
1510
1511
1512
1513
1514
        if not isinstance(tensor_1, list):
            tensor_1 = tensor_1.cpu().tolist()
        if not isinstance(tensor_2, list):
            tensor_2 = tensor_2.cpu().tolist()

        in_order = len(tensor_1) <= len(tensor_2)
1515
1516
1517
1518
        longer = tensor_2 if in_order else tensor_1
        shorter = tensor_1 if in_order else tensor_2

        flag = False
1519
1520
        chunk_size = len(shorter)
        for chunk_idx in range(len(longer) - chunk_size + 1):
1521
            subseq = longer[chunk_idx : chunk_idx + chunk_size]
1522
            if subseq == shorter:
1523
1524
1525
1526
1527
                flag = True
                break

        self.assertTrue(flag)

1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631

@require_torch
class UtilsFunctionsTest(unittest.TestCase):

    # tests whether the top_k_top_p function behaves as expected
    def test_top_k_top_p_filtering(self):
        logits = torch.tensor(
            [
                [
                    8.2220991,  # 3rd highest value; idx. 0
                    -0.5620044,
                    5.23229752,
                    4.0386393,
                    -6.8798378,
                    -0.54785802,
                    -3.2012153,
                    2.92777176,
                    1.88171953,
                    7.35341276,
                    8.43207833,  # 2nd highest value; idx. 10
                    -9.85711836,
                    -5.96209236,
                    -1.13039161,
                    -7.1115294,
                    -0.8369633,
                    -5.3186408,
                    7.06427407,
                    0.81369344,
                    -0.82023817,
                    -5.9179796,
                    0.58813443,
                    -6.99778438,
                    4.71551189,
                    -0.18771637,
                    7.44020759,  # 4th highest value; idx. 25
                    9.38450987,  # 1st highest value; idx. 26
                    2.12662941,
                    -9.32562038,
                    2.35652522,
                ],  # cummulative prob of 4 highest values <= 0.6
                [
                    0.58425518,
                    4.53139238,
                    -5.57510464,
                    -6.28030699,
                    -7.19529503,
                    -4.02122551,
                    1.39337037,
                    -6.06707057,
                    1.59480517,
                    -9.643119,
                    0.03907799,
                    0.67231762,
                    -8.88206726,
                    6.27115922,  # 4th highest value; idx. 13
                    2.28520723,
                    4.82767506,
                    4.30421368,
                    8.8275313,  # 2nd highest value; idx. 17
                    5.44029958,
                    -4.4735794,
                    7.38579536,  # 3rd highest value; idx. 20
                    -2.91051663,
                    2.61946077,
                    -2.5674762,
                    -9.48959302,
                    -4.02922645,
                    -1.35416918,
                    9.67702323,  # 1st highest value; idx. 27
                    -5.89478553,
                    1.85370467,
                ],  # cummulative prob of 4 highest values <= 0.6
            ],
            dtype=torch.float,
            device=torch_device,
        )

        non_inf_expected_idx = torch.tensor(
            [[0, 0], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 20], [1, 27]],
            dtype=torch.long,
            device=torch_device,
        )  # expected non filtered idx as noted above

        non_inf_expected_output = torch.tensor(
            [
                8.2221,
                8.4321,
                7.4402,
                9.3845,
                6.2712,
                8.8275,
                7.3858,
                9.6770,
            ],  # expected non filtered values as noted above
            dtype=torch.float,
            device=torch_device,
        )

        output = top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
        non_inf_output = output[output != -float("inf")].to(device=torch_device)
        non_inf_idx = (output != -float("inf")).nonzero().to(device=torch_device)

        self.assertTrue(torch.allclose(non_inf_expected_output, non_inf_output, atol=1e-12))
        self.assertTrue(torch.all(torch.eq(non_inf_expected_idx, non_inf_idx)))
1632

1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
    # tests whether the function uses filter_value instead of default -inf
    def test_top_k_top_p_filtering_with_filter_value(self):
        logits = torch.tensor(
            [
                [
                    1,
                    1,
                    1,
                    0.99,  # get filtered by top-p filtering
                    0.98,  # get filtered by top-k filtering
                ]
            ],
            dtype=torch.float,
            device=torch_device,
        )

        expected_output = torch.tensor(
            [[1, 1, 1, 0, 0]],
            dtype=torch.float,
            device=torch_device,
        )

        output = top_k_top_p_filtering(logits, top_k=4, top_p=0.5, filter_value=0.0)

        self.assertTrue(torch.allclose(expected_output, output, atol=1e-12))

1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673

@require_torch
class GenerationIntegrationTests(unittest.TestCase):
    @slow
    def test_diverse_beam_search(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood.
        The celebrity couple announced the arrival of their son, Silas Randall Timberlake, in statements to People.
        "Silas was the middle name of Timberlake's maternal grandfather Bill Bomar, who died in 2012, while Randall is the musician's own middle name, as well as his father's first," People reports.
        The couple announced the pregnancy in January, with an Instagram post. It is the first baby for both."""

        bart_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
        bart_model = BartForConditionalGeneration.from_pretrained("facebook/bart-large-cnn").to(torch_device)
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        outputs = bart_model.generate(
1674
1675
1676
1677
1678
1679
            input_ids,
            num_beams=4,
            num_return_sequences=2,
            num_beam_groups=4,
            diversity_penalty=2.0,
            remove_invalid_values=True,
1680
1681
1682
1683
1684
1685
1686
        )

        generated_text = bart_tokenizer.batch_decode(outputs, skip_special_tokens=True)

        self.assertListEqual(
            generated_text,
            [
Sylvain Gugger's avatar
Sylvain Gugger committed
1687
1688
1689
1690
1691
1692
                "The couple announced the birth of their son, Silas Randall Timberlake, in a statement. Silas was the"
                " middle name of Timberlake's maternal grandfather Bill Bomar. Randall is the musician's own middle"
                " name, as well as his father's first. It is the first baby for both of them.",
                "Justin Timberlake and Jessica Biel have a son. The baby is named Silas Randall Timberlake. It is the"
                " first child for both. The couple announced the pregnancy in January. The name Silas is the middle"
                " name of Timberlake's maternal grandfather. It's also his own middle name.",
1693
1694
            ],
        )
1695
1696
1697

    def test_max_length_backward_compat_greedy(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1698
1699
1700
1701
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1702
1703
1704
1705
1706
1707
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        max_length = 20
        input_ids = input_ids.expand(2, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
        input_ids = bart_model._prepare_decoder_input_ids_for_generation(
1708
            input_ids.shape[0],
1709
1710
1711
1712
            decoder_start_token_id=bart_model.config.decoder_start_token_id,
            bos_token_id=bart_model.config.bos_token_id,
        )

1713
1714
1715
1716
1717
1718
1719
1720
        with self.assertWarns(UserWarning):
            bart_model.greedy_search(
                input_ids,
                max_length=max_length,
                pad_token_id=bart_model.config.pad_token_id,
                eos_token_id=bart_model.config.eos_token_id,
                **model_kwargs,
            )
1721
1722
1723

    def test_max_length_backward_compat_sample(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1724
1725
1726
1727
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1728
1729
1730
1731
1732
1733
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        max_length = 20
        input_ids = input_ids.expand(2, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
        input_ids = bart_model._prepare_decoder_input_ids_for_generation(
1734
            input_ids.shape[0],
1735
1736
1737
            decoder_start_token_id=bart_model.config.decoder_start_token_id,
            bos_token_id=bart_model.config.bos_token_id,
        )
1738
        with torch.no_grad():
1739
1740
1741
1742
1743
1744
1745
1746
            with self.assertWarns(UserWarning):
                bart_model.sample(
                    input_ids,
                    max_length=max_length,
                    pad_token_id=bart_model.config.pad_token_id,
                    eos_token_id=bart_model.config.eos_token_id,
                    **model_kwargs,
                )
1747
1748
1749

    def test_max_length_backward_compat_beam_search(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1750
1751
1752
1753
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1754
1755
1756
1757
1758
1759
1760
1761
1762
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        batch_size = 1
        max_length = 20
        num_beams = 2

        input_ids = input_ids.expand(2, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
        input_ids = bart_model._prepare_decoder_input_ids_for_generation(
1763
            input_ids.shape[0],
1764
1765
1766
1767
1768
1769
1770
1771
1772
            decoder_start_token_id=bart_model.config.decoder_start_token_id,
            bos_token_id=bart_model.config.bos_token_id,
        )

        beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=num_beams,
            device=torch_device,
        )
1773
1774
1775
1776
        with self.assertWarns(UserWarning):
            _ = bart_model.beam_search(
                input_ids, num_beams=num_beams, max_length=max_length, beam_scorer=beam_scorer, **model_kwargs
            )
1777
1778
1779

    def test_max_length_backward_compat_group_beam_search(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1780
1781
1782
1783
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        batch_size = 1
        max_length = 20
        num_beams = 6
        num_beam_groups = 3
        num_return_sequences = num_beams * batch_size

        input_ids = input_ids.expand(6, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
        input_ids = bart_model._prepare_decoder_input_ids_for_generation(
1795
            input_ids.shape[0],
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
            decoder_start_token_id=bart_model.config.decoder_start_token_id,
            bos_token_id=bart_model.config.bos_token_id,
        )

        diverse_beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=num_beams,
            device=torch_device,
            num_beam_hyps_to_keep=num_return_sequences,
            num_beam_groups=num_beam_groups,
        )
1807
1808
1809
1810
        with self.assertWarns(UserWarning):
            bart_model.group_beam_search(
                input_ids, diverse_beam_scorer, num_beams=num_beams, max_length=max_length, **model_kwargs
            )
1811
1812
1813

    def test_max_length_warning_if_different(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1814
1815
1816
1817
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        batch_size = 1

        max_length = 20
        num_beams = 6
        num_beam_groups = 3
        num_return_sequences = num_beams * batch_size
        stopping_criteria_max_length = 18
        stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=stopping_criteria_max_length)])

        # Greedy
        input_ids = input_ids.expand(6, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
        input_ids = bart_model._prepare_decoder_input_ids_for_generation(
1833
            input_ids.shape[0],
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
            decoder_start_token_id=bart_model.config.decoder_start_token_id,
            bos_token_id=bart_model.config.bos_token_id,
        )

        with self.assertWarns(UserWarning):
            bart_model.greedy_search(
                input_ids,
                max_length=max_length,
                pad_token_id=bart_model.config.pad_token_id,
                stopping_criteria=stopping_criteria,
                eos_token_id=bart_model.config.eos_token_id,
                **model_kwargs,
            )

        # Sample
        with self.assertWarns(UserWarning):
1850
1851
1852
1853
1854
1855
1856
1857
1858
            with torch.no_grad():
                bart_model.sample(
                    input_ids,
                    max_length=max_length,
                    stopping_criteria=stopping_criteria,
                    pad_token_id=bart_model.config.pad_token_id,
                    eos_token_id=bart_model.config.eos_token_id,
                    **model_kwargs,
                )
1859
1860
1861
1862
1863
1864
1865
1866

        # Beam
        beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=num_beams,
            device=torch_device,
        )
        with self.assertWarns(UserWarning):
1867
1868
1869
1870
1871
1872
1873
1874
1875
            with torch.no_grad():
                bart_model.beam_search(
                    input_ids,
                    num_beams=num_beams,
                    stopping_criteria=stopping_criteria,
                    max_length=max_length,
                    beam_scorer=beam_scorer,
                    **model_kwargs,
                )
1876
1877
1878
1879
1880
1881
1882
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893

        # Grouped beam search
        diverse_beam_scorer = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=num_beams,
            device=torch_device,
            num_beam_hyps_to_keep=num_return_sequences,
            num_beam_groups=num_beam_groups,
        )
        with self.assertWarns(UserWarning):
            bart_model.group_beam_search(
                input_ids,
                diverse_beam_scorer,
                stopping_criteria=stopping_criteria,
                num_beams=num_beams,
                max_length=max_length,
                **model_kwargs,
            )
1894
1895
1896

    def test_beam_search_warning_if_max_length_is_passed(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
1897
1898
1899
1900
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
1901
1902
1903
1904
1905
1906
1907
1908

        batch_size = 1
        num_beams = 3

        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        input_ids = input_ids.expand(num_beams, -1)
        model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})

1909
1910
1911
        # pretend decoder_input_ids correspond to first encoder input id
        decoder_input_ids = input_ids[:, :1]

1912
1913
1914
1915
1916
1917
1918
1919
1920
1921
1922
1923
        stopping_criteria_max_length = 18
        stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=stopping_criteria_max_length)])

        with self.assertWarns(UserWarning):
            beam_scorer = BeamSearchScorer(
                batch_size=batch_size,
                num_beams=num_beams,
                device=torch_device,
                max_length=10,
            )

        generated_ids = bart_model.beam_search(
1924
            decoder_input_ids,
1925
1926
1927
1928
1929
1930
1931
1932
1933
1934
1935
1936
1937
            num_beams=num_beams,
            stopping_criteria=stopping_criteria,
            beam_scorer=beam_scorer,
            **model_kwargs,
        )

        beam_scorer_no_max_len = BeamSearchScorer(
            batch_size=batch_size,
            num_beams=num_beams,
            device=torch_device,
        )

        generated_ids_no_max_len = bart_model.beam_search(
1938
            decoder_input_ids,
1939
1940
1941
1942
1943
1944
1945
1946
            num_beams=num_beams,
            stopping_criteria=stopping_criteria,
            beam_scorer=beam_scorer_no_max_len,
            **model_kwargs,
        )

        # BeamSearchScorer max_length should not influence "real" max_length
        self.assertEqual(generated_ids.tolist(), generated_ids_no_max_len.tolist())
1947

1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
    def test_custom_stopping_criteria_overload_error(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
        bart_tokenizer = BartTokenizer.from_pretrained("sshleifer/bart-tiny-random")
        bart_model = BartForConditionalGeneration.from_pretrained("sshleifer/bart-tiny-random").to(torch_device)

        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        stopping_criteria = StoppingCriteriaList()
        stopping_criteria.append(MaxLengthCriteria(max_length=42))
        with self.assertRaises(ValueError):
            bart_model.generate(input_ids, stopping_criteria=stopping_criteria)
        with self.assertRaises(ValueError):
            bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=32)

    def test_custom_stopping_criteria(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
        bart_tokenizer = BartTokenizer.from_pretrained("sshleifer/bart-tiny-random")
        bart_model = BartForConditionalGeneration.from_pretrained("sshleifer/bart-tiny-random").to(torch_device)
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        class DummyCriteria(StoppingCriteria):
            def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
                return input_ids.shape[-1] >= 20

        stopping_criteria = StoppingCriteriaList()
        stopping_criteria.append(DummyCriteria())

        self.assertEqual(
            list(bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=22).shape),
            [1, 20],
        )
        self.assertEqual(
            list(bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=18).shape),
            [1, 18],
        )

1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
    def test_stop_sequence_stopping_criteria(self):

        prompt = """Hello I believe in"""
        generator = pipeline("text-generation", model="hf-internal-testing/tiny-random-bart")
        output = generator(prompt)
        self.assertEqual(
            output,
            [
                {
                    "generated_text": (
                        "Hello I believe in in in number number number number number number number number number"
                    )
                }
            ],
        )

        output = generator(prompt, stop_sequence=" number")
        self.assertEqual(output, [{"generated_text": "Hello I believe in in in number"}])

2002
2003
2004
    def test_custom_logits_processor(self):
        bart_tokenizer = BartTokenizer.from_pretrained("sshleifer/bart-tiny-random")
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
2005
2006
2007
        bart_model = BartForConditionalGeneration.from_pretrained("sshleifer/bart-tiny-random", min_length=1).to(
            torch_device
        )
2008
2009
2010
2011
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        logits_processor = LogitsProcessorList()
        logits_processor.append(MinLengthLogitsProcessor(min_length=10, eos_token_id=0))
2012
        # it should not be allowed to both define `min_length` via config and `logits_processor` list
2013
2014
2015
2016
2017
2018
        with self.assertRaises(ValueError):
            bart_model.generate(input_ids, logits_processor=logits_processor)

        bart_model.config.min_length = None
        bart_model.generate(input_ids, logits_processor=logits_processor)

2019
    def test_max_new_tokens_encoder_decoder(self):
2020
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
2021
2022
2023
2024
        bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
            torch_device
        )
2025
2026
        input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

2027
        self.assertEqual(list(input_ids.shape), [1, 29])
2028
2029

        max_new_tokens = 3
2030
        bart_model.config.max_length = 20
2031
        bart_model.config.eos_token_id = None
2032
2033

        # Encoder decoder call
2034
2035
2036
2037
2038
2039
        outputs = bart_model.generate(input_ids, max_new_tokens=max_new_tokens)
        # 1 BOS + 3 new tokens
        self.assertEqual(list(outputs.shape), [1, 4])

        # Decoder only call
        outputs = bart_model.generate(decoder_input_ids=input_ids, max_new_tokens=max_new_tokens)
2040
2041
        # 29 + 3 new tokens
        self.assertEqual(list(outputs.shape), [1, 32])
2042

2043
2044
2045
2046
2047
2048
        # Encoder decoder call > 20
        outputs = bart_model.generate(max_new_tokens=max_new_tokens + 20)

        # 1 BOS + 20 + 3 new tokens
        self.assertEqual(list(outputs.shape), [1, 24])

2049
2050
        # max_new_tokens and max_length serve the same purpose and must not be used together.
        with self.assertRaises(ValueError):
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
            bart_model.generate(decoder_input_ids=input_ids, max_new_tokens=10, max_length=20)

    def test_max_new_tokens_decoder_only(self):
        article = """Justin Timberlake."""
        gpt2_tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        gpt2_model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
        input_ids = gpt2_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)

        self.assertEqual(list(input_ids.shape), [1, 9])

        max_new_tokens = 3
        gpt2_model.config.max_length = 20

        # call < 20
        outputs = gpt2_model.generate(input_ids, max_new_tokens=max_new_tokens)

        # 9 input_ids + 3 new tokens
        self.assertEqual(list(outputs.shape), [1, 12])

        # call > 20
        outputs = gpt2_model.generate(max_new_tokens=max_new_tokens + 20)

        # 1 BOS token + 23 new tokens
        self.assertEqual(list(outputs.shape), [1, 24])

2076
2077
        # max_new_tokens and max_length serve the same purpose and must not be used together.
        with self.assertRaises(ValueError):
2078
            gpt2_model.generate(input_ids=input_ids, max_new_tokens=10, max_length=20)
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094

    def test_encoder_decoder_generate_with_inputs_embeds(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart", max_length=5).to(
            torch_device
        )
        model.config.eos_token_id = None
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        inputs_embeds = model.get_input_embeddings()(input_ids)

        output_sequences = model.generate(inputs_embeds=inputs_embeds)

        # make sure model generated correctly until `max_length`
        self.assertEqual(output_sequences.shape, (1, 5))

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
    def test_encoder_decoder_generate_attention_mask(self):
        articles = ["Timberlake", "Jessica Biel, welcome to parenthood among other things"]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        # need extrem generation values here to force this test
        # to fail when `attention_mask` is not correctly treated in generate
        model = BartForConditionalGeneration.from_pretrained(
            "hf-internal-testing/tiny-random-bart", max_length=50, num_beams=5, num_return_sequences=5
        ).to(torch_device)

        model.config.eos_token_id = None
        input_ids = tokenizer(articles[0], return_tensors="pt").input_ids.to(torch_device)
        input_ids_batched = tokenizer(articles, padding=True, return_tensors="pt").input_ids.to(torch_device)

        output_sequences_batched = model.generate(
            input_ids=input_ids_batched, return_dict_in_generate=True, output_scores=True
        )
        output_sequences = model.generate(input_ids=input_ids, return_dict_in_generate=True, output_scores=True)

        batched_out = output_sequences_batched.sequences_scores
        out = output_sequences.sequences_scores

        diff = (batched_out[:5].sum() - out.sum()).abs()

        self.assertTrue(diff < 1e-4)

2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
    def test_decoder_generate_with_inputs_embeds(self):
        article = """I need input_ids to generate"""
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2", max_length=5).to(torch_device)
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        inputs_embeds = model.get_input_embeddings()(input_ids)

        # cannot generate from `inputs_embeds` for decoder only
        with self.assertRaises(ValueError):
            model.generate(inputs_embeds=inputs_embeds)
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141

    def test_generate_input_ids_as_kwarg(self):
        article = """I need input_ids to generate"""
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2", max_length=15).to(torch_device)
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        output_sequences_kwargs = model.generate(input_ids=input_ids).cpu()
        output_sequences = model.generate(input_ids).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (1, 15))

2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
    def test_generate_non_nlp_input_ids_as_kwarg(self):
        model = ImageGPTForCausalImageModeling.from_pretrained(
            "hf-internal-testing/tiny-random-imagegpt", max_length=10
        ).to(torch_device)
        input_ids = ids_tensor((3, 5), vocab_size=10)

        output_sequences_kwargs = model.generate(input_ids=input_ids).cpu()
        output_sequences = model.generate(input_ids).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (3, 10))

2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
    def test_generate_input_ids_as_encoder_kwarg(self):
        article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart", max_length=5).to(
            torch_device
        )
        model.config.eos_token_id = None
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        output_sequences_kwargs = model.generate(input_ids=input_ids).cpu()
        output_sequences = model.generate(input_ids).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (1, 5))

    def test_generate_inputs_and_encoder_kwargs(self):
        article = """I need input_ids to generate"""
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2", max_length=10).to(torch_device)
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        with self.assertRaises(ValueError):
            model.generate(input_ids, input_ids=input_ids)

    def test_generate_too_many_encoder_kwargs(self):
        article = """I need input_ids to generate"""
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2", max_length=10).to(torch_device)
        input_ids = tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
        with self.assertRaises(ValueError):
2182
            model.generate(input_ids=input_ids, inputs_embeds=input_ids)
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212

    def test_generate_input_values_as_encoder_kwarg(self):
        input_values = floats_tensor((2, 250))
        model = SpeechEncoderDecoderModel.from_pretrained("hf-internal-testing/tiny-random-speech-encoder-decoder")
        model = model.to(torch_device)
        output_sequences_kwargs = model.generate(input_values=input_values, max_length=5).cpu()
        output_sequences = model.generate(input_values, max_length=5).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (2, 5))

    def test_generate_input_features_as_encoder_kwarg(self):
        input_features = floats_tensor((3, 20, 24))
        model = Speech2TextForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-speech_to_text")
        model = model.to(torch_device)
        output_sequences_kwargs = model.generate(input_features=input_features, max_length=5).cpu()
        output_sequences = model.generate(input_features, max_length=5).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (3, 5))

    def test_generate_pixel_values_as_encoder_kwarg(self):
        pixel_values = floats_tensor((2, 3, 30, 30))
        model = VisionEncoderDecoderModel.from_pretrained("hf-internal-testing/tiny-random-vision-encoder-decoder")
        model = model.to(torch_device)
        output_sequences_kwargs = model.generate(pixel_values=pixel_values, max_length=5).cpu()
        output_sequences = model.generate(pixel_values, max_length=5).cpu()

        self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
        self.assertEqual(output_sequences.shape, (2, 5))
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228

    def test_generate_encoder_outputs_attention_mask(self):
        input_values = floats_tensor((2, 250)).to(torch_device)
        attention_mask = torch.ones_like(input_values)
        model = SpeechEncoderDecoderModel.from_pretrained("hf-internal-testing/tiny-random-speech-encoder-decoder")
        model = model.to(torch_device)

        encoder = model.get_encoder()

        encoder_outputs = encoder(input_values)

        output_sequences_no_mask = model.generate(encoder_outputs=encoder_outputs).cpu()
        output_sequences_with_mask = model.generate(encoder_outputs=encoder_outputs, attention_mask=attention_mask)
        output_sequences_with_mask = output_sequences_with_mask.cpu()

        self.assertListEqual(output_sequences_no_mask.tolist(), output_sequences_with_mask.tolist())
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
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
2296
2297
2298
2299
2300
2301
2302
2303
2304
2305
2306
2307
2308
2309
2310
2311
2312
2313
2314
2315
2316
2317
2318
2319
2320
2321
2322
2323
2324
2325
2326
2327
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338
2339
2340
2341
2342
2343
2344
2345
2346
2347
2348
2349
2350
2351
2352
2353
2354
2355
2356
2357
2358
2359
2360
2361
2362
2363
2364
2365
2366
2367
2368
2369
2370
2371
2372

    def test_transition_scores_beam_search_encoder_decoder(self):
        articles = [
            "Justin Timberlake and Jessica Biel, welcome to parenthood.",
            "Michael Phelps is arguably the most decorated Olympian of all time.",
        ]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained(
            "hf-internal-testing/tiny-random-bart",
            max_length=10,
            num_beams=4,
            num_return_sequences=2,
            eos_token_id=None,
            return_dict_in_generate=True,
            output_scores=True,
            length_penalty=0.0,
        )
        model = model.to(torch_device)

        input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
        outputs = model.generate(input_ids=input_ids)

        transition_scores = model.compute_transition_beam_scores(
            outputs.sequences, outputs.scores, outputs.beam_indices
        )
        transition_scores_sum = transition_scores.sum(-1)

        self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))

    def test_transition_scores_beam_search_encoder_decoder_with_eos(self):
        articles = [
            "Justin Timberlake and Jessica Biel, welcome to parenthood.",
            "Michael Phelps is arguably the most decorated Olympian of all time.",
        ]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained(
            "hf-internal-testing/tiny-random-bart",
            max_length=10,
            num_beams=4,
            num_return_sequences=2,
            return_dict_in_generate=True,
            output_scores=True,
            length_penalty=0.0,
        )
        model = model.to(torch_device)

        input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
        outputs = model.generate(input_ids=input_ids)

        transition_scores = model.compute_transition_beam_scores(
            outputs.sequences, outputs.scores, outputs.beam_indices
        )
        transition_scores_sum = transition_scores.sum(-1)

        self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))

    def test_transition_scores_beam_search_decoder_only(self):
        articles = [
            "Justin Timberlake",
            "Michael Phelps",
        ]
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        tokenizer.pad_token = tokenizer.eos_token

        model = GPT2LMHeadModel.from_pretrained(
            "hf-internal-testing/tiny-random-gpt2",
            max_length=10,
            num_beams=4,
            num_return_sequences=2,
            pad_token_id=tokenizer.eos_token_id,
            eos_token_id=None,
            return_dict_in_generate=True,
            output_scores=True,
            length_penalty=0.0,
        )
        model = model.to(torch_device)

        input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
        outputs = model.generate(input_ids=input_ids)

        transition_scores = model.compute_transition_beam_scores(
            outputs.sequences, outputs.scores, outputs.beam_indices
        )
        transition_scores_sum = transition_scores.sum(-1)

        self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))

    def test_transition_scores_beam_sample_encoder_decoder(self):
        articles = [
            "Justin Timberlake and Jessica Biel, welcome to parenthood.",
            "Michael Phelps is arguably the most decorated Olympian of all time.",
        ]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained(
            "hf-internal-testing/tiny-random-bart",
            do_sample=True,
            max_length=10,
            num_beams=4,
            num_return_sequences=2,
            eos_token_id=None,
            return_dict_in_generate=True,
            output_scores=True,
            length_penalty=0.0,
        )
        model = model.to(torch_device)

        input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
        outputs = model.generate(input_ids=input_ids)

        transition_scores = model.compute_transition_beam_scores(
            outputs.sequences, outputs.scores, outputs.beam_indices
        )
        transition_scores_sum = transition_scores.sum(-1)

        self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))

    def test_transition_scores_group_beam_search_encoder_decoder(self):
        articles = [
            "Justin Timberlake and Jessica Biel, welcome to parenthood.",
            "Michael Phelps is arguably the most decorated Olympian of all time.",
        ]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained(
            "hf-internal-testing/tiny-random-bart",
            max_length=10,
            num_beams=2,
            num_beam_groups=2,
            num_return_sequences=2,
            eos_token_id=None,
            return_dict_in_generate=True,
            output_scores=True,
            length_penalty=0.0,
        )
        model = model.to(torch_device)

        input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
        outputs = model.generate(input_ids=input_ids)

        transition_scores = model.compute_transition_beam_scores(
            outputs.sequences, outputs.scores, outputs.beam_indices
        )
        transition_scores_sum = transition_scores.sum(-1)

        self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
2373

2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385
2386
2387
2388
2389
2390
2391
2392
2393
2394
2395
2396
2397
2398
2399
2400
2401
2402
2403
2404
2405
2406
2407
2408
2409
2410
2411
2412
2413
2414
2415
2416
2417
2418
2419
2420
2421
2422
2423
2424
2425
2426
2427
2428
2429
2430
2431
2432
2433
2434
2435
2436
2437
2438
2439
2440
2441
2442
2443
2444
2445
2446
2447
2448
2449
2450
2451
2452
2453
2454
2455
2456
2457
2458
2459
2460
2461
    @slow
    def test_transition_scores_early_stopping(self):
        # This is an aggressive test that makes sure that `beam_search's`
        # transition scores are computed correctly for varying `num_return_sequences`,
        # `num_beams` and `batch_size > 1`
        # 2 x input_ids for "question: How are you? \n context: I had a long day, "
        input_ids = torch.tensor(2 * [[822, 10, 571, 33, 25, 58, 2625, 10, 27, 141, 3, 9, 307, 239, 6, 1]]).to(
            torch_device
        )

        model = AutoModelForSeq2SeqLM.from_pretrained("t5-small").to(torch_device)

        result = model.generate(
            input_ids,
            max_length=10,
            return_dict_in_generate=True,
            output_scores=True,
            forced_eos_token_id=model.config.eos_token_id,
            num_beams=4,
            do_sample=False,
            num_return_sequences=3,
            length_penalty=0.0,
        )

        transition_scores = model.compute_transition_beam_scores(
            sequences=result.sequences, scores=result.scores, beam_indices=result.beam_indices
        )

        sum_transition_scores = torch.sum(transition_scores, dim=1)

        self.assertListEqual(sum_transition_scores.cpu().tolist(), result.sequences_scores.cpu().tolist())

    def test_log_scores_sample_decoder_only(self):
        articles = ["I need input_ids to generate", "Short and"]
        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
        tokenizer.padding_side = "left"
        tokenizer.pad_token = tokenizer.eos_token

        model = GPT2LMHeadModel.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)

        inputs = tokenizer(articles, return_tensors="pt", padding=True).to(torch_device)

        result = model.generate(
            **inputs,
            max_length=15,
            return_dict_in_generate=True,
            do_sample=False,
            output_scores=True,
        )

        # decoder-only starts generating from `input_ids`
        begin_generation = inputs.input_ids.shape[-1]

        gen_sequences = result.sequences[:, begin_generation:]
        probs = torch.stack(result.scores, dim=1).softmax(-1)

        gen_probs = torch.gather(probs, 2, gen_sequences[:, :, None]).squeeze(-1)
        expected_probs = torch.tensor([[0.0014, 0.0015], [0.0014, 0.0014]])

        self.assertTrue(torch.allclose(gen_probs.cpu(), expected_probs, atol=1e-3))

    def test_log_scores_sample_encoder_decoder(self):
        articles = ["I need input_ids to generate", "Short and"]
        tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
        model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(torch_device)

        inputs = tokenizer(articles, return_tensors="pt", padding=True).to(torch_device)

        result = model.generate(
            **inputs,
            max_length=3,
            return_dict_in_generate=True,
            do_sample=False,
            num_beams=1,
            output_scores=True,
        )

        # encoder-decoder has one decoder_start_token_id by default
        begin_generation = 1

        gen_sequences = result.sequences[:, begin_generation:]
        probs = torch.stack(result.scores, dim=1).softmax(-1)

        gen_probs = torch.gather(probs, 2, gen_sequences[:, :, None]).squeeze(-1)
        expected_probs = torch.tensor([[0.0013, 1.0000], [0.0013, 1.0000]])

        self.assertTrue(torch.allclose(gen_probs.cpu(), expected_probs, atol=1e-3))

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
2501
2502
2503
    @slow
    def test_beam_search_example_integration(self):
        # exactly the example provided in the docstrings of beam search, which previously
        # failed after directly copying from it. Refer to PR #15555
        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)
        outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)

        self.assertListEqual(outputs, ["Wie alt bist du?"])

2504
2505
    @slow
    def test_constrained_beam_search(self):
2506
2507
        model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
2508

2509
2510
        force_tokens = tokenizer("scared", add_prefix_space=True, add_special_tokens=False).input_ids
        force_tokens_2 = tokenizer("big weapons", add_prefix_space=True, add_special_tokens=False).input_ids
2511
2512
2513
2514
2515
2516
2517
2518
2519
2520
2521
2522
2523
2524
2525
2526
2527
2528
2529
2530
2531
2532
2533
2534
2535

        constraints = [
            PhrasalConstraint(force_tokens),
            PhrasalConstraint(force_tokens_2),
        ]

        starting_text = ["The soldiers were not prepared and"]

        input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)

        outputs = model.generate(
            input_ids,
            constraints=constraints,
            num_beams=10,
            num_return_sequences=1,
            no_repeat_ngram_size=1,
            max_length=30,
            remove_invalid_values=True,
        )

        generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)

        self.assertListEqual(
            generated_text,
            [
2536
2537
                "The soldiers were not prepared and didn't know what to do. They had no idea how they would react if"
                " the enemy attacked them, big weapons scared"
2538
2539
2540
            ],
        )

2541
2542
    @slow
    def test_constrained_beam_search_mixed(self):
2543
2544
        model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
2545
2546
2547
2548
2549
2550
2551
2552
2553
2554
2555
2556
2557
2558
2559
2560
2561
2562
2563
2564
2565
2566
2567
2568
2569
2570
2571
2572
2573
2574

        force_phrase = tokenizer("scared", add_prefix_space=True, add_special_tokens=False).input_ids
        flexible_phrases = tokenizer(
            ["scream", "screams", "screaming", "screamed"], add_prefix_space=True, add_special_tokens=False
        ).input_ids

        constraints = [
            PhrasalConstraint(force_phrase),
            DisjunctiveConstraint(flexible_phrases),
        ]

        starting_text = ["The soldiers", "The child"]

        input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)

        outputs = model.generate(
            input_ids,
            constraints=constraints,
            num_beams=10,
            num_return_sequences=1,
            no_repeat_ngram_size=1,
            # max_length=20,
            remove_invalid_values=True,
        )

        generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)

        self.assertListEqual(
            generated_text,
            [
2575
2576
2577
                "The soldiers, who had been stationed at the base for more than a year before being evacuated"
                " screaming scared",
                "The child was taken to a local hospital where he died.\n 'I don't think screaming scared",
2578
2579
2580
2581
2582
            ],
        )

    @slow
    def test_constrained_beam_search_mixed_mixin(self):
2583
2584
        model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
2585
2586
2587
2588
2589
2590
2591
2592
2593
2594
2595
2596
2597
2598
2599
2600
2601
2602
2603
2604
2605
2606
2607
2608
2609
2610
2611

        force_word = "scared"
        force_flexible = ["scream", "screams", "screaming", "screamed"]

        force_words_ids = [
            tokenizer([force_word], add_prefix_space=True, add_special_tokens=False).input_ids,
            tokenizer(force_flexible, add_prefix_space=True, add_special_tokens=False).input_ids,
        ]

        starting_text = ["The soldiers", "The child"]

        input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)

        outputs = model.generate(
            input_ids,
            force_words_ids=force_words_ids,
            num_beams=10,
            num_return_sequences=1,
            no_repeat_ngram_size=1,
            remove_invalid_values=True,
        )

        generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)

        self.assertListEqual(
            generated_text,
            [
2612
2613
2614
                "The soldiers, who had been stationed at the base for more than a year before being evacuated"
                " screaming scared",
                "The child was taken to a local hospital where he died.\n 'I don't think screaming scared",
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
            ],
        )

    @slow
    def test_constrained_beam_search_example_translation_mixin(self):
        tokenizer = AutoTokenizer.from_pretrained("t5-base")
        model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")

        encoder_input_str = "translate English to German: How old are you?"
        force_words = ["sind"]

        input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids
        force_words_ids = tokenizer(force_words, add_special_tokens=False).input_ids

        outputs = model.generate(
            input_ids,
            force_words_ids=force_words_ids,
            num_beams=10,
            num_return_sequences=1,
            no_repeat_ngram_size=1,
            remove_invalid_values=True,
        )

        outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)

2640
        self.assertListEqual(outputs, ["Wie alt sind Sie?"])
2641

2642
2643
2644
2645
2646
2647
2648
2649
2650
2651
2652
2653
2654
2655
2656
2657
2658
2659
2660
2661
2662
2663
2664
2665
2666
2667
2668
2669
2670
2671
2672
2673
2674
2675
2676
2677
2678
2679
2680
2681
2682
2683
    @slow
    def test_constrained_beam_search_example_integration(self):
        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 5 beams
        num_beams = 5
        # 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 = "sind"
        constraint_token_ids = tokenizer.encode(constraint_str)[:-1]  # 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
        )
        outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)

2684
        self.assertListEqual(outputs, ["Wie alt sind Sie?"])
2685
2686

    def test_constrained_beam_search_mixin_type_checks(self):
2687
2688
        tokenizer = AutoTokenizer.from_pretrained("patrickvonplaten/t5-tiny-random")
        model = AutoModelForSeq2SeqLM.from_pretrained("patrickvonplaten/t5-tiny-random")
2689
2690
2691
2692
2693
2694
2695
2696
2697
2698
2699
2700
2701
2702
2703
2704
2705
2706
2707
2708
2709
2710
2711
2712
2713
2714
2715
2716
2717
2718
2719
2720
2721
2722
2723
2724

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

        with self.assertRaises(ValueError):
            force_words = ["sind"]
            force_words_ids = tokenizer(force_words, return_tensors="pt").input_ids
            model.generate(
                input_ids,
                force_words_ids=force_words_ids,
                num_beams=10,
                num_return_sequences=1,
                no_repeat_ngram_size=1,
                remove_invalid_values=True,
            )

        with self.assertRaises(ValueError):
            force_words = ["sind"]
            force_words_ids = [tokenizer(force_words, return_tensors="pt").input_ids]
            model.generate(
                input_ids,
                force_words_ids=force_words_ids,
                num_beams=10,
                num_return_sequences=1,
                no_repeat_ngram_size=1,
                remove_invalid_values=True,
            )

        with self.assertRaises(ValueError):
            model.generate(input_ids, force_words_ids=[])

        with self.assertRaises(ValueError):
            model.generate(input_ids, force_words_ids=[[-1]])

        with self.assertRaises(ValueError):
            model.generate(input_ids, force_words_ids=[[[-1]]])
2725
2726

    def test_validate_generation_inputs(self):
2727
2728
        tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
        model = AutoModelForSeq2SeqLM.from_pretrained("hf-internal-testing/tiny-random-t5")
2729
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740

        encoder_input_str = "Hello world"
        input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids

        # typos are quickly detected (the correct argument is `do_sample`)
        with self.assertRaisesRegex(ValueError, "do_samples"):
            model.generate(input_ids, do_samples=True)

        # arbitrary arguments that will not be used anywhere are also not accepted
        with self.assertRaisesRegex(ValueError, "foo"):
            fake_model_kwargs = {"foo": "bar"}
            model.generate(input_ids, **fake_model_kwargs)