File size: 36,123 Bytes
4c28b8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34783fc
4c28b8d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Team All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for recipe-generation.
"""
# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.

import logging
import os
import random
import sys
import time
from dataclasses import dataclass, field
from functools import partial
from pathlib import Path
from typing import Callable, Optional

import datasets
import nltk  # Here to have a nice missing dependency error message early on
import numpy as np
from datasets import Dataset, load_dataset, load_metric
from tqdm import tqdm

import jax
import jax.numpy as jnp
import optax
import transformers
from filelock import FileLock
from flax import jax_utils, traverse_util
from flax.jax_utils import unreplicate
from flax.training import train_state
from flax.training.common_utils import get_metrics, onehot, shard, shard_prng_key
from transformers import (
    CONFIG_MAPPING,
    FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
    AutoConfig,
    AutoTokenizer,
    FlaxAutoModelForSeq2SeqLM,
    HfArgumentParser,
    TrainingArguments,
    is_tensorboard_available,
)
from transformers.file_utils import is_offline_mode

logger = logging.getLogger(__name__)

try:
    nltk.data.find("tokenizers/punkt")
except (LookupError, OSError):
    if is_offline_mode():
        raise LookupError(
            "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
        )
    with FileLock(".lock") as lock:
        nltk.download("punkt", quiet=True)

MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
    """

    model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": "The model checkpoint for weights initialization."
                    "Don't set if you want to train a model from scratch."
        },
    )
    model_type: Optional[str] = field(
        default=None,
        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"}
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    dtype: Optional[str] = field(
        default="float32",
        metadata={
            "help": "Floating-point format in which the model weights should be initialized and trained. Choose one of `[float32, float16, bfloat16]`."
        },
    )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    text_column: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the column in the datasets containing the inputs (for generation)."},
    )
    target_column: Optional[str] = field(
        default=None,
        metadata={"help": "The name of the column in the datasets containing the targets (for generation)."},
    )
    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
    validation_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
    )
    max_source_length: Optional[int] = field(
        default=128,
        metadata={
            "help": "The maximum total input sequence length after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded."
        },
    )
    max_target_length: Optional[int] = field(
        default=1024,
        metadata={
            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded."
        },
    )
    val_max_target_length: Optional[int] = field(
        default=None,
        metadata={
            "help": "The maximum total sequence length for validation target text after tokenization. Sequences longer "
                    "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
                    "This argument is also used to override the `max_length` param of `model.generate`, which is used "
                    "during evaluation."
        },
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
                    "value if set."
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                    "value if set."
        },
    )
    max_predict_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                    "value if set."
        },
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    source_prefix: Optional[str] = field(
        default=None, metadata={"help": "A prefix to add before every source text (useful for T5 models)."}
    )
    predict_with_generate: bool = field(
        default=False, metadata={"help": "Whether to use generate to calculate generative metrics (ROUGE, BLEU)."}
    )
    num_beams: Optional[int] = field(
        default=None,
        metadata={
            "help": "Number of beams to use for evaluation. This argument will be passed to `model.generate`, "
                    "which is used during evaluation."
        },
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    extra_tokens: str = field(
        default=None,
        metadata={"help": "A text list of extra tokens separated by `,` that you want to add to the vocab."},
    )
    special_tokens: str = field(
        default=None,
        metadata={"help": "A list of special tokens separated by `,` that you want to add to the vocab."},
    )
    prediction_debug: bool = field(
        default=False,
        metadata={
            "help": "Whether to show some examples of the model prediction"
        },
    )

    def __post_init__(self):
        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
            raise ValueError("Need either a dataset name or a training/validation file.")
        else:
            if self.train_file is not None:
                extension = self.train_file.split(".")[-1]
                assert extension in ["csv", "json"], "`train_file` should be a csv or a json file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension in ["csv", "json"], "`validation_file` should be a csv or a json file."
        if self.val_max_target_length is None:
            self.val_max_target_length = self.max_target_length


class TrainState(train_state.TrainState):
    dropout_rng: jnp.ndarray

    def replicate(self):
        return jax_utils.replicate(self).replace(dropout_rng=shard_prng_key(self.dropout_rng))


def data_loader(rng: jax.random.PRNGKey, dataset: Dataset, batch_size: int, shuffle: bool = False):
    """
    Returns batches of size `batch_size` from truncated `dataset`, sharded over all local devices.
    Shuffle batches if `shuffle` is `True`.
    """
    steps_per_epoch = len(dataset) // batch_size

    if shuffle:
        batch_idx = jax.random.permutation(rng, len(dataset))
    else:
        batch_idx = jnp.arange(len(dataset))

    batch_idx = batch_idx[: steps_per_epoch * batch_size]  # Skip incomplete batch.
    batch_idx = batch_idx.reshape((steps_per_epoch, batch_size))

    for idx in batch_idx:
        batch = dataset[idx]
        batch = {k: jnp.array(v) for k, v in batch.items()}

        batch = shard(batch)

        yield batch


def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
    summary_writer.scalar("train_time", train_time, step)

    train_metrics = get_metrics(train_metrics)
    for key, vals in train_metrics.items():
        tag = f"train_{key}"
        for i, val in enumerate(vals):
            summary_writer.scalar(tag, val, step - len(vals) + i + 1)

    for metric_name, value in eval_metrics.items():
        summary_writer.scalar(f"eval_{metric_name}", value, step)


def create_learning_rate_fn(
        train_ds_size: int, train_batch_size: int, num_train_epochs: int, num_warmup_steps: int, learning_rate: float
) -> Callable[[int], jnp.array]:
    """Returns a linear warmup, linear_decay learning rate function."""
    steps_per_epoch = train_ds_size // train_batch_size
    num_train_steps = steps_per_epoch * num_train_epochs
    warmup_fn = optax.linear_schedule(init_value=0.0, end_value=learning_rate, transition_steps=num_warmup_steps)
    decay_fn = optax.linear_schedule(
        init_value=learning_rate, end_value=0, transition_steps=num_train_steps - num_warmup_steps
    )
    schedule_fn = optax.join_schedules(schedules=[warmup_fn, decay_fn], boundaries=[num_warmup_steps])
    return schedule_fn


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    if (
            os.path.exists(training_args.output_dir)
            and os.listdir(training_args.output_dir)
            and training_args.do_train
            and not training_args.overwrite_output_dir
    ):
        raise ValueError(
            f"Output directory ({training_args.output_dir}) already exists and is not empty."
            "Use --overwrite_output_dir to overcome."
        )

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    # Setup logging, we only want one process per machine to log things on the screen.
    logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR)
    if jax.process_index() == 0:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # Set the verbosity to info of the Transformers logger (on main process only):
    logger.info(f"Training/evaluation parameters {training_args}")

    # Get the datasets: you can either provide your own CSV/JSON training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files this script will use the first column for the full texts and the second column for the
    # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
    #
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            keep_in_memory=False
        )
    else:
        data_files = {}

        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
        if data_args.test_file is not None:
            data_files["test"] = data_args.test_file
            extension = data_args.test_file.split(".")[-1]

        dataset = load_dataset(
            extension,
            data_files=data_files,
            delimiter="\t",
            cache_dir=model_args.cache_dir
        )
    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer

    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer
        )
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
        )

    if data_args.extra_tokens and isinstance(data_args.extra_tokens, str):
        extra_tokens = list(data_args.extra_tokens.split(","))
        if len(extra_tokens) > 0:
            logger.info(f"*** Adding extra tokens: {extra_tokens} ***")
            tokenizer.add_tokens(extra_tokens, special_tokens=False)

    if data_args.special_tokens and isinstance(data_args.special_tokens, str):
        special_tokens = list(data_args.special_tokens.split(","))
        if len(special_tokens) > 0:
            logger.info(f"*** Adding special tokens: {special_tokens} ***")
            tokenizer.add_tokens(special_tokens, special_tokens=True)

    if model_args.model_name_or_path:
        model = FlaxAutoModelForSeq2SeqLM.from_pretrained(
            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )
    else:
        model = FlaxAutoModelForSeq2SeqLM.from_config(
            config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )

    model.resize_token_embeddings(len(tokenizer))
    if model.config.decoder_start_token_id is None:
        raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")

    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    if training_args.do_train:
        column_names = dataset["train"].column_names
    elif training_args.do_eval:
        column_names = dataset["validation"].column_names
    elif training_args.do_predict:
        column_names = dataset["test"].column_names
    else:
        logger.info("There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`.")
        return

    # Get the column names for input/target.
    if data_args.text_column is None:
        text_column = column_names[0]
    else:
        text_column = data_args.text_column
        if text_column not in column_names:
            raise ValueError(
                f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
            )
    if data_args.target_column is None:
        target_column = column_names[1]
    else:
        target_column = data_args.target_column
        if target_column not in column_names:
            raise ValueError(
                f"--target_column' value '{data_args.target_column}' needs to be one of: {', '.join(column_names)}"
            )

    # Temporarily set max_target_length for training.
    max_target_length = data_args.max_target_length

    # In Flax, for seq2seq models we need to pass `decoder_input_ids`
    # as the Flax models don't accept `labels`, we need to prepare the decoder_input_ids here
    # for that dynamically import the `shift_tokens_right` function from the model file
    model_module = __import__(model.__module__, fromlist=["shift_tokens_tight"])
    shift_tokens_right_fn = getattr(model_module, "shift_tokens_right")

    # Setting padding="max_length" as we need fixed length inputs for jitted functions
    def preprocess_function(examples):
        inputs = examples[text_column]
        targets = examples[target_column]
        inputs = [prefix + inp for inp in inputs]
        model_inputs = tokenizer(
            inputs, max_length=data_args.max_source_length, padding="max_length", truncation=True, return_tensors="np"
        )

        # Setup the tokenizer for targets
        with tokenizer.as_target_tokenizer():
            labels = tokenizer(
                targets, max_length=max_target_length, padding="max_length", truncation=True, return_tensors="np"
            )

        model_inputs["labels"] = labels["input_ids"]
        decoder_input_ids = shift_tokens_right_fn(
            jnp.array(labels["input_ids"]), config.pad_token_id, config.decoder_start_token_id
        )
        model_inputs["decoder_input_ids"] = np.asarray(decoder_input_ids)

        # We need decoder_attention_mask so we can ignore pad tokens from loss
        model_inputs["decoder_attention_mask"] = labels["attention_mask"]

        return model_inputs

    if training_args.do_train:
        if "train" not in dataset:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = dataset["train"]
        if data_args.max_train_samples is not None:
            train_dataset = train_dataset.select(range(data_args.max_train_samples))
        train_dataset = train_dataset.map(
            preprocess_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on train dataset",
        )

    if training_args.do_eval:
        max_target_length = data_args.val_max_target_length
        if "validation" not in dataset:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = dataset["validation"]
        if data_args.max_eval_samples is not None:
            eval_dataset = eval_dataset.select(range(data_args.max_eval_samples))
        eval_dataset = eval_dataset.map(
            preprocess_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on validation dataset",
        )

    if training_args.do_predict:
        max_target_length = data_args.val_max_target_length
        if "test" not in dataset:
            raise ValueError("--do_predict requires a test dataset")
        predict_dataset = dataset["test"]
        if data_args.max_predict_samples is not None:
            predict_dataset = predict_dataset.select(range(data_args.max_predict_samples))
        predict_dataset = predict_dataset.map(
            preprocess_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
            desc="Running tokenizer on prediction dataset",
        )

    # Metrics
    bleu = load_metric("sacrebleu")
    wer = load_metric("wer")

    def postprocess_text(preds, labels):
        preds = [pred.strip() for pred in preds]
        labels_bleu = [[label.strip()] for label in labels]
        labels_wer = [label.strip() for label in labels]

        return preds, [labels_bleu, labels_wer]

    def compute_metrics(preds, labels):
        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

        # Some simple post-processing
        decoded_preds, [decoded_labels_bleu, decoded_labels_wer] = postprocess_text(decoded_preds, decoded_labels)

        if data_args.prediction_debug:
            for index in random.sample(range(len(decoded_labels)), 3):
                logger.info(f'reference: "{decoded_labels[index]}"')
                logger.info(f'predicted: "{decoded_preds[index]}"')
                logger.info('---')

        result = {}

        try:
            result_blue = bleu.compute(predictions=decoded_preds, references=decoded_labels_wer)
            result_blue = result_blue["score"]
        except Exception as e:
            logger.info(f'Error occurred during bleu {e}')
            result_blue = 0.0 * 100

        try:
            result_wer = wer.compute(predictions=decoded_preds, references=decoded_labels_wer)
            result_wer = result_wer * 100
        except Exception as e:
            logger.info(f'Error occurred during wer {e}')
            result_wer = 1.0 * 100

        result["blue"] = result_blue
        result["wer"] = result_wer

        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
        result["gen_len"] = np.mean(prediction_lens)
        result = {k: round(v, 4) for k, v in result.items()}
        return result

    # Enable tensorboard only on the master node
    has_tensorboard = is_tensorboard_available()
    if has_tensorboard and jax.process_index() == 0:
        try:
            from flax.metrics.tensorboard import SummaryWriter

            summary_writer = SummaryWriter(log_dir=Path(training_args.output_dir))
        except ImportError as ie:
            has_tensorboard = False
            logger.warning(
                f"Unable to display metrics through TensorBoard because some package are not installed: {ie}"
            )
    else:
        logger.warning(
            "Unable to display metrics through TensorBoard because the package is not installed: "
            "Please run pip install tensorboard to enable."
        )

    # Initialize our training
    rng = jax.random.PRNGKey(training_args.seed)
    rng, dropout_rng = jax.random.split(rng)

    # Store some constant
    num_epochs = int(training_args.num_train_epochs)
    train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()
    eval_batch_size = int(training_args.per_device_eval_batch_size) * jax.device_count()
    steps_per_epoch = len(train_dataset) // train_batch_size
    total_train_steps = steps_per_epoch * num_epochs

    # Create learning rate schedule
    linear_decay_lr_schedule_fn = create_learning_rate_fn(
        len(train_dataset),
        train_batch_size,
        training_args.num_train_epochs,
        training_args.warmup_steps,
        training_args.learning_rate,
    )

    # We use Optax's "masking" functionality to not apply weight decay
    # to bias and LayerNorm scale parameters. decay_mask_fn returns a
    # mask boolean with the same structure as the parameters.
    # The mask is True for parameters that should be decayed.
    # Note that this mask is specifically adapted for FlaxBart.
    # For FlaxT5, one should correct the layer norm parameter naming
    # accordingly - see `run_t5_mlm_flax.py` e.g.

    if any(x in model_args.model_name_or_path for x in ["t5", "mt5", "byt5"]):
        def decay_mask_fn(params):
            flat_params = traverse_util.flatten_dict(params)
            layer_norm_params = [
                (name, "scale") for name in ["layer_norm", "final_layer_norm"]
            ]
            flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_params) for path in flat_params}
            return traverse_util.unflatten_dict(flat_mask)
    else:
        def decay_mask_fn(params):
            flat_params = traverse_util.flatten_dict(params)
            layer_norm_params = [
                (name, "scale") for name in ["self_attn_layer_norm", "layernorm_embedding", "final_layer_norm"]
            ]
            flat_mask = {path: (path[-1] != "bias" and path[-2:] not in layer_norm_params) for path in flat_params}
            return traverse_util.unflatten_dict(flat_mask)

    # create adam optimizer
    adamw = optax.adamw(
        learning_rate=linear_decay_lr_schedule_fn,
        b1=training_args.adam_beta1,
        b2=training_args.adam_beta2,
        eps=training_args.adam_epsilon,
        weight_decay=training_args.weight_decay,
        mask=decay_mask_fn,
    )

    # Setup train state
    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)

    # label smoothed cross entropy
    def loss_fn(logits, labels, padding_mask, label_smoothing_factor=0.0):
        """
        The label smoothing implementation is adapted from Flax's official example:
        https://github.com/google/flax/blob/87a211135c6a377c8f29048a1cac3840e38b9da4/examples/wmt/train.py#L104
        """
        vocab_size = logits.shape[-1]
        confidence = 1.0 - label_smoothing_factor
        low_confidence = (1.0 - confidence) / (vocab_size - 1)
        normalizing_constant = -(
                confidence * jnp.log(confidence) + (vocab_size - 1) * low_confidence * jnp.log(low_confidence + 1e-20)
        )
        soft_labels = onehot(labels, vocab_size, on_value=confidence, off_value=low_confidence)

        loss = optax.softmax_cross_entropy(logits, soft_labels)
        loss = loss - normalizing_constant

        # ignore padded tokens from loss
        loss = loss * padding_mask
        loss = loss.sum() / padding_mask.sum()
        return loss

    # Define gradient update step fn
    def train_step(state, batch, label_smoothing_factor=0.0):
        dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)

        def compute_loss(params):
            labels = batch.pop("labels")
            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
            loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor)
            return loss

        grad_fn = jax.value_and_grad(compute_loss)
        loss, grad = grad_fn(state.params)
        grad = jax.lax.pmean(grad, "batch")

        new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng)

        metrics = {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}
        metrics = jax.lax.pmean(metrics, axis_name="batch")

        return new_state, metrics

    # Define eval fn
    def eval_step(params, batch, label_smoothing_factor=0.0):
        labels = batch.pop("labels")
        logits = model(**batch, params=params, train=False)[0]
        loss = loss_fn(logits, labels, batch["decoder_attention_mask"], label_smoothing_factor)

        # summarize metrics
        metrics = {"loss": loss}
        metrics = jax.lax.pmean(metrics, axis_name="batch")
        return metrics

    # Define generation function
    max_length = (
        data_args.val_max_target_length if data_args.val_max_target_length is not None else model.config.max_length
    )
    num_beams = data_args.num_beams if data_args.num_beams is not None else model.config.num_beams
    gen_kwargs = {"max_length": max_length, "num_beams": num_beams}

    def generate_step(params, batch):
        model.params = params
        output_ids = model.generate(batch["input_ids"], attention_mask=batch["attention_mask"], **gen_kwargs)
        return output_ids.sequences

    # Create parallel version of the train and eval step
    p_train_step = jax.pmap(
        partial(train_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch", donate_argnums=(0,)
    )
    p_eval_step = jax.pmap(partial(eval_step, label_smoothing_factor=training_args.label_smoothing_factor), "batch")
    p_generate_step = jax.pmap(generate_step, "batch")

    # Replicate the train state on each device
    state = state.replicate()

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {num_epochs}")
    logger.info(f"  Instantaneous batch size per device = {training_args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel & distributed) = {train_batch_size}")
    logger.info(f"  Total optimization steps = {total_train_steps}")

    train_time = 0
    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
    for epoch in epochs:
        # ======================== Training ================================
        train_start = time.time()

        # Create sampling rng
        rng, input_rng = jax.random.split(rng)
        train_metrics = []

        # Generate an epoch by shuffling sampling indices from the train dataset
        train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True)
        steps_per_epoch = len(train_dataset) // train_batch_size
        # train
        for _ in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False):
            batch = next(train_loader)
            state, train_metric = p_train_step(state, batch)
            train_metrics.append(train_metric)

        train_time += time.time() - train_start

        train_metric = unreplicate(train_metric)

        epochs.write(
            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
        )

        # ======================== Evaluating ==============================
        eval_metrics = []
        eval_preds = []
        eval_labels = []

        eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size)
        eval_steps = len(eval_dataset) // eval_batch_size
        for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False):
            # Model forward
            batch = next(eval_loader)
            labels = batch["labels"]

            metrics = p_eval_step(state.params, batch)
            eval_metrics.append(metrics)

            # generation
            if data_args.predict_with_generate:
                generated_ids = p_generate_step(state.params, batch)
                eval_preds.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
                eval_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))

        # normalize eval metrics
        eval_metrics = get_metrics(eval_metrics)
        eval_metrics = jax.tree_map(jnp.mean, eval_metrics)

        # compute ROUGE metrics
        rouge_desc = ""
        if data_args.predict_with_generate:
            rouge_metrics = compute_metrics(eval_preds, eval_labels)
            eval_metrics.update(rouge_metrics)
            rouge_desc = " ".join([f"Eval {key}: {value} |" for key, value in rouge_metrics.items()])

        # Print metrics and update progress bar
        desc = f"Epoch... ({epoch + 1}/{num_epochs} | Eval Loss: {eval_metrics['loss']} | {rouge_desc})"
        epochs.write(desc)
        epochs.desc = desc

        # Save metrics
        if has_tensorboard and jax.process_index() == 0:
            cur_step = epoch * (len(train_dataset) // train_batch_size)
            write_metric(summary_writer, train_metrics, eval_metrics, train_time, cur_step)

    # ======================== Prediction loop ==============================
    if training_args.do_predict:
        logger.info("*** Predict ***")

        pred_metrics = []
        pred_generations = []
        pred_labels = []

        pred_loader = data_loader(input_rng, predict_dataset, eval_batch_size)
        pred_steps = len(predict_dataset) // eval_batch_size
        for _ in tqdm(range(pred_steps), desc="Predicting...", position=2, leave=False):
            # Model forward
            batch = next(pred_loader)
            labels = batch["labels"]

            metrics = p_eval_step(state.params, batch)
            pred_metrics.append(metrics)

            # generation
            if data_args.predict_with_generate:
                generated_ids = p_generate_step(state.params, batch)
                pred_generations.extend(jax.device_get(generated_ids.reshape(-1, gen_kwargs["max_length"])))
                pred_labels.extend(jax.device_get(labels.reshape(-1, labels.shape[-1])))

        # normalize prediction metrics
        pred_metrics = get_metrics(pred_metrics)
        pred_metrics = jax.tree_map(jnp.mean, pred_metrics)

        # compute ROUGE metrics
        rouge_desc = ""
        if data_args.predict_with_generate:
            rouge_metrics = compute_metrics(pred_generations, pred_labels)
            pred_metrics.update(rouge_metrics)
            rouge_desc = " ".join([f"Predict {key}: {value} |" for key, value in rouge_metrics.items()])

        # Print metrics
        desc = f"Predict Loss: {pred_metrics['loss']} | {rouge_desc})"
        logger.info(desc)

        # save checkpoint after each epoch and push checkpoint to the hub
        if jax.process_index() == 0:
            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
            model.save_pretrained(
                training_args.output_dir,
                params=params,
                push_to_hub=training_args.push_to_hub,
                commit_message=f"Saving weights and logs of epoch {epoch + 1}",
            )


if __name__ == "__main__":
    main()