just starting training

events.out.tfevents.1627341143.t1v-n-55481057-w-0.9623.3.v2

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0df2b578828b327778dc02eb8ca788d5e99b1cf8acaad0b93f22cdec22aab58a
+size 147065

run.sh

@@ -0,0 +1,19 @@
+./run_t5_mlm_flax_streaming.py\
+    --model_name_or_path="./" \
+    --output_dir="./" \
+    --config_name="./" \
+    --tokenizer_name="./" \
+    --dataset_name="pere/nb_nn_balanced_shuffled" \
+    --max_seq_length="512" \
+    --per_device_train_batch_size="32" \
+    --per_device_eval_batch_size="32" \
+    --learning_rate="0.005" \
+    --weight_decay="0.001" \
+    --warmup_steps="2000" \
+    --overwrite_output_dir \
+    --logging_steps="500" \
+    --save_steps="2500" \
+    --eval_steps="2500"\
+    --push_to_hub \
+    --preprocessing_num_workers 96 \
+    --adafactor 

run_t5_mlm_flax_streaming.py

@@ -0,0 +1,751 @@
+#!/usr/bin/env python3
+# 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.
+"""
+Pretraining with T5-like span-masked language modeling on a streaming dataset.
+Here is the full list of checkpoints on the hub that can be pretrained by this script:
+https://huggingface.co/models?filter=t5
+"""
+import logging
+import os
+import sys
+import time
+from collections import defaultdict
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Dict, Optional
+
+import datasets
+import numpy as np
+from datasets import load_dataset
+from tqdm import tqdm
+
+import flax
+import jax
+import jax.numpy as jnp
+import optax
+from flax import jax_utils, traverse_util
+from flax.training import train_state
+from flax.training.common_utils import get_metrics, onehot, shard
+from transformers import (
+    CONFIG_MAPPING,
+    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
+    BatchEncoding,
+    FlaxT5ForConditionalGeneration,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    T5Config,
+    T5TokenizerFast,
+    TrainingArguments,
+    is_tensorboard_available,
+    set_seed,
+)
+from transformers.models.t5.modeling_flax_t5 import shift_tokens_right
+
+if datasets.__version__ <= "1.8.0":
+    raise ValueError("Make sure to upgrade `datasets` to a version >= 1.9.0 to use dataset streaming")
+
+MODEL_CONFIG_CLASSES = list(FLAX_MODEL_FOR_MASKED_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]`."
+        },
+    )
+    auth_token: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Auth token for private repositories on the Huggingface Hub"
+        }
+    )
+
+
+@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)."}
+    )
+    max_seq_length: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "The maximum total input sequence length after tokenization and masking. Sequences longer than this will be truncated. Default to the max input length of the model."
+        },
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    mlm_probability: float = field(
+        default=0.15, metadata={"help": "Ratio of tokens to mask for span masked language modeling loss"}
+    )
+    mean_noise_span_length: float = field(
+        default=3.0,
+        metadata={"help": "Mean span length of masked tokens"},
+    )
+    text_column_name: str = field(
+        default="text", metadata={"help": "The name of the column to retrieve the training text."}
+    )
+    shuffle_buffer_size: int = field(
+        default=10000, metadata={"help": "The number of examples to pre-load for shuffling."}
+    )
+    num_train_steps: int = field(default=50000, metadata={"help": "The number of training steps."})
+    num_eval_samples: int = field(default=50000, metadata={"help": "The number of samples to be used for evaluation"})
+
+    def __post_init__(self):
+        if self.dataset_name is None:
+            raise ValueError("Need a dataset name for streaming.")
+
+
+def compute_input_and_target_lengths(inputs_length, noise_density, mean_noise_span_length):
+    """This function is copy of `random_spans_helper <https://github.com/google-research/text-to-text-transfer-transformer/blob/84f8bcc14b5f2c03de51bd3587609ba8f6bbd1cd/t5/data/preprocessors.py#L2466>`__ .
+    Training parameters to avoid padding with random_spans_noise_mask.
+    When training a model with random_spans_noise_mask, we would like to set the other
+    training hyperparmeters in a way that avoids padding.
+    This function helps us compute these hyperparameters.
+    We assume that each noise span in the input is replaced by extra_tokens_per_span_inputs sentinel tokens,
+    and each non-noise span in the targets is replaced by extra_tokens_per_span_targets sentinel tokens.
+    This function tells us the required number of tokens in the raw example (for split_tokens())
+    as well as the length of the encoded targets. Note that this function assumes
+    the inputs and targets will have EOS appended and includes that in the reported length.
+    Args:
+        inputs_length: an integer - desired length of the tokenized inputs sequence
+        noise_density: a float
+        mean_noise_span_length: a float
+    Returns:
+        tokens_length: length of original text in tokens
+        targets_length: an integer - length in tokens of encoded targets sequence
+    """
+
+    def _tokens_length_to_inputs_length_targets_length(tokens_length):
+        num_noise_tokens = int(round(tokens_length * noise_density))
+        num_nonnoise_tokens = tokens_length - num_noise_tokens
+        num_noise_spans = int(round(num_noise_tokens / mean_noise_span_length))
+        # inputs contain all nonnoise tokens, sentinels for all noise spans
+        # and one EOS token.
+        _input_length = num_nonnoise_tokens + num_noise_spans + 1
+        _output_length = num_noise_tokens + num_noise_spans + 1
+        return _input_length, _output_length
+
+    tokens_length = inputs_length
+
+    while _tokens_length_to_inputs_length_targets_length(tokens_length + 1)[0] <= inputs_length:
+        tokens_length += 1
+
+    inputs_length, targets_length = _tokens_length_to_inputs_length_targets_length(tokens_length)
+
+    # minor hack to get the targets length to be equal to inputs length
+    # which is more likely to have been set to a nice round number.
+    if noise_density == 0.5 and targets_length > inputs_length:
+        tokens_length -= 1
+        targets_length -= 1
+    return tokens_length, targets_length
+
+
+@flax.struct.dataclass
+class FlaxDataCollatorForT5MLM:
+    """
+    Data collator used for T5 span-masked language modeling.
+    It is made sure that after masking the inputs are of length `data_args.max_seq_length` and targets are also of fixed length.
+    For more information on how T5 span-masked language modeling works, one can take a look
+    at the `official paper <https://arxiv.org/pdf/1910.10683.pdf>`__
+    or the `official code for preprocessing <https://github.com/google-research/text-to-text-transfer-transformer/blob/master/t5/data/preprocessors.py>`__ .
+    Args:
+        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
+            The tokenizer used for encoding the data.
+        noise_density (:obj:`float`):
+            The probability with which to (randomly) mask tokens in the input.
+        mean_noise_span_length (:obj:`float`):
+            The average span length of the masked tokens.
+        input_length (:obj:`int`):
+            The expected input length after masking.
+        target_length (:obj:`int`):
+            The expected target length after masking.
+        pad_token_id: (:obj:`int`):
+            The pad token id of the model
+        decoder_start_token_id: (:obj:`int):
+            The decoder start token id of the model
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    noise_density: float
+    mean_noise_span_length: float
+    input_length: int
+    target_length: int
+    pad_token_id: int
+    decoder_start_token_id: int
+
+    def __call__(self, examples: Dict[str, np.ndarray]) -> BatchEncoding:
+        
+        batch = BatchEncoding(
+            {k: np.array(examples[k]) for k in examples.keys()}
+        )
+        input_ids = batch['input_ids']
+        batch_size, expandend_input_length = input_ids.shape
+
+        mask_indices = np.asarray([self.random_spans_noise_mask(expandend_input_length) for i in range(batch_size)])
+        labels_mask = ~mask_indices
+
+        input_ids_sentinel = self.create_sentinel_ids(mask_indices.astype(np.int8))
+        labels_sentinel = self.create_sentinel_ids(labels_mask.astype(np.int8))
+
+        batch["input_ids"] = self.filter_input_ids(input_ids, input_ids_sentinel)
+        batch["labels"] = self.filter_input_ids(input_ids, labels_sentinel)
+
+        if batch["input_ids"].shape[-1] != self.input_length:
+            raise ValueError(
+                f"`input_ids` are incorrectly preprocessed. `input_ids` length is {batch['input_ids'].shape[-1]}, but should be {self.target_length}."
+            )
+
+        if batch["labels"].shape[-1] != self.target_length:
+            raise ValueError(
+                f"`labels` are incorrectly preprocessed. `labels` length is {batch['labels'].shape[-1]}, but should be {self.target_length}."
+            )
+
+        # to check that tokens are correctly proprocessed, one can run `self.tokenizer.batch_decode(input_ids)` and `self.tokenizer.batch_decode(labels)` here...
+        batch["decoder_input_ids"] = shift_tokens_right(
+            batch["labels"], self.pad_token_id, self.decoder_start_token_id
+        )
+
+        return batch
+
+    def create_sentinel_ids(self, mask_indices):
+        """
+        Sentinel ids creation given the indices that should be masked.
+        The start indices of each mask are replaced by the sentinel ids in increasing
+        order. Consecutive mask indices to be deleted are replaced with `-1`.
+        """
+        start_indices = mask_indices - np.roll(mask_indices, 1, axis=-1) * mask_indices
+        start_indices[:, 0] = mask_indices[:, 0]
+
+        sentinel_ids = np.where(start_indices != 0, np.cumsum(start_indices, axis=-1), start_indices)
+        sentinel_ids = np.where(sentinel_ids != 0, (sentinel_ids + self.tokenizer.vocab_size - 1), 0)
+        sentinel_ids -= mask_indices - start_indices
+
+        return sentinel_ids
+
+    def filter_input_ids(self, input_ids, sentinel_ids):
+        """
+        Puts sentinel mask on `input_ids` and fuse consecutive mask tokens into a single mask token by deleting.
+        This will reduce the sequence length from `expanded_inputs_length` to `input_length`.
+        """
+        batch_size = input_ids.shape[0]
+
+        input_ids_full = np.where(sentinel_ids != 0, sentinel_ids, input_ids)
+        input_ids = input_ids_full[input_ids_full > 0].reshape((batch_size, -1))
+        input_ids = np.concatenate(
+            [input_ids, np.full((batch_size, 1), self.tokenizer.eos_token_id, dtype=np.int32)], axis=-1
+        )
+        return input_ids
+
+    def random_spans_noise_mask(self, length):
+
+        """This function is copy of `random_spans_helper <https://github.com/google-research/text-to-text-transfer-transformer/blob/84f8bcc14b5f2c03de51bd3587609ba8f6bbd1cd/t5/data/preprocessors.py#L2682>`__ .
+        Noise mask consisting of random spans of noise tokens.
+        The number of noise tokens and the number of noise spans and non-noise spans
+        are determined deterministically as follows:
+        num_noise_tokens = round(length * noise_density)
+        num_nonnoise_spans = num_noise_spans = round(num_noise_tokens / mean_noise_span_length)
+        Spans alternate between non-noise and noise, beginning with non-noise.
+        Subject to the above restrictions, all masks are equally likely.
+        Args:
+            length: an int32 scalar (length of the incoming token sequence)
+            noise_density: a float - approximate density of output mask
+            mean_noise_span_length: a number
+        Returns:
+            a boolean tensor with shape [length]
+        """
+
+        orig_length = length
+
+        num_noise_tokens = int(np.round(length * self.noise_density))
+        # avoid degeneracy by ensuring positive numbers of noise and nonnoise tokens.
+        num_noise_tokens = min(max(num_noise_tokens, 1), length - 1)
+        num_noise_spans = int(np.round(num_noise_tokens / self.mean_noise_span_length))
+
+        # avoid degeneracy by ensuring positive number of noise spans
+        num_noise_spans = max(num_noise_spans, 1)
+        num_nonnoise_tokens = length - num_noise_tokens
+
+        # pick the lengths of the noise spans and the non-noise spans
+        def _random_segmentation(num_items, num_segments):
+            """Partition a sequence of items randomly into non-empty segments.
+            Args:
+                num_items: an integer scalar > 0
+                num_segments: an integer scalar in [1, num_items]
+            Returns:
+                a Tensor with shape [num_segments] containing positive integers that add
+                up to num_items
+            """
+            mask_indices = np.arange(num_items - 1) < (num_segments - 1)
+            np.random.shuffle(mask_indices)
+            first_in_segment = np.pad(mask_indices, [[1, 0]])
+            segment_id = np.cumsum(first_in_segment)
+            segment_length = np.asarray(jax.ops.segment_sum(np.ones_like(segment_id), segment_id))
+            return segment_length
+
+        noise_span_lengths = _random_segmentation(num_noise_tokens, num_noise_spans)
+        nonnoise_span_lengths = _random_segmentation(num_nonnoise_tokens, num_noise_spans)
+
+        interleaved_span_lengths = np.reshape(
+            np.stack([nonnoise_span_lengths, noise_span_lengths], axis=1), [num_noise_spans * 2]
+        )
+        span_starts = np.cumsum(interleaved_span_lengths)[:-1]
+        span_start_indicator = np.zeros((length,), dtype=np.int8)
+        span_start_indicator[span_starts] = True
+        span_num = np.cumsum(span_start_indicator)
+        is_noise = np.equal(span_num % 2, 1)
+
+        return is_noise[:orig_length]
+
+
+def generate_batch_splits(samples_idx: jnp.ndarray, batch_size: int) -> jnp.ndarray:
+    num_samples = len(samples_idx)
+    samples_to_remove = num_samples % batch_size
+
+    if samples_to_remove != 0:
+        samples_idx = samples_idx[:-samples_to_remove]
+    sections_split = num_samples // batch_size
+    batch_idx = np.split(samples_idx, sections_split)
+    return batch_idx
+
+
+def advance_iter_and_group_samples(train_iterator, num_samples, max_seq_length):
+    """
+    The training iterator is advanced so that after groupifying the samples,
+    `num_samples` of length `max_seq_length` are returned.
+    """
+    num_total_tokens = max_seq_length * num_samples
+    samples = defaultdict(list)
+
+    i = 0
+    while i < num_total_tokens:
+        tokenized_samples = next(train_iterator)
+        i += len(tokenized_samples["input_ids"])
+
+        # concatenate tokenized samples to list
+        samples = {k: samples[k] + tokenized_samples[k] for k in tokenized_samples.keys()}
+
+    # Concatenated tokens are split to lists of length `max_seq_length`.
+    # Note that remainedr of % max_seq_length are thrown away.
+    def group_texts(examples):
+        result = {
+            k: [t[i : i + max_seq_length] for i in range(0, num_total_tokens, max_seq_length)]
+            for k, t in examples.items()
+        }
+        return result
+
+    grouped_samples = group_texts(samples)
+    return grouped_samples
+
+
+def write_train_metric(summary_writer, train_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)
+
+
+def write_eval_metric(summary_writer, eval_metrics, step):
+    for metric_name, value in eval_metrics.items():
+        summary_writer.scalar(f"eval_{metric_name}", value, step)
+
+
+if __name__ == "__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."
+        )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        level="INFO",
+        datefmt="[%X]",
+    )
+
+    # Log on each process the small summary:
+    logger = logging.getLogger(__name__)
+    #logger.warning(
+    #    f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+    #    + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    #)
+
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    logger.info(f"Training/evaluation parameters {training_args}")
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # Get the datasets: you can either provide your own CSV/JSON/TXT 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 column called 'text' or the first column if no column called
+    # 'text' is found. You can easily tweak this behavior (see below).
+    if data_args.dataset_name is not None:
+        # Downloading and loading a dataset from the hub.
+        datasets = load_dataset(
+            data_args.dataset_name, 
+            data_args.dataset_config_name, 
+            cache_dir=model_args.cache_dir,
+            streaming=True,
+            split="train"
+        )
+
+    # 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.tokenizer_name:
+        tokenizer = T5TokenizerFast.from_pretrained(
+            model_args.tokenizer_name, 
+            cache_dir=model_args.cache_dir, 
+            use_fast=model_args.use_fast_tokenizer,
+            use_auth_token=model_args.auth_token
+        )
+    elif model_args.model_name_or_path:
+        tokenizer = T5TokenizerFast.from_pretrained(
+            model_args.model_name_or_path, 
+            cache_dir=model_args.cache_dir, 
+            use_fast=model_args.use_fast_tokenizer,
+            use_auth_token=model_args.auth_token
+        )
+    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 model_args.config_name:
+        config = T5Config.from_pretrained(
+            model_args.config_name, cache_dir=model_args.cache_dir, vocab_size=len(tokenizer)
+        )
+    elif model_args.model_name_or_path:
+        config = T5Config.from_pretrained(
+            model_args.model_name_or_path, cache_dir=model_args.cache_dir, vocab_size=len(tokenizer)
+        )
+    else:
+        config = CONFIG_MAPPING[model_args.model_type]()
+        logger.warning("You are instantiating a new config instance from scratch.")
+
+    # Preprocessing the datasets.
+    # First we tokenize all the texts.
+
+    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)
+
+    # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
+    # Since we make sure that all sequences are of the same length, no attention_mask is needed.
+    def tokenize_function(examples):
+        return tokenizer(examples[data_args.text_column_name], return_attention_mask=False)
+
+    tokenized_datasets = datasets.map(
+        tokenize_function,
+        batched=True
+    )
+
+    # T5-like span masked language modeling will fuse consecutively masked tokens to a single sentinel token.
+    # To ensure that the input length is `max_seq_length`, we need to increase the maximum length
+    # according to `mlm_probability` and `mean_noise_span_length`. We can also define the label length accordingly.
+    expanded_inputs_length, targets_length = compute_input_and_target_lengths(
+        inputs_length=max_seq_length,
+        noise_density=data_args.mlm_probability,
+        mean_noise_span_length=data_args.mean_noise_span_length,
+    )
+
+    shuffle_seed = training_args.seed
+    tokenized_datasets = tokenized_datasets.shuffle(buffer_size=data_args.shuffle_buffer_size, seed=shuffle_seed)
+
+    # 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)
+    dropout_rngs = jax.random.split(rng, jax.local_device_count())
+
+    model = FlaxT5ForConditionalGeneration(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))
+
+    # Data collator
+    # This one will take care of randomly masking the tokens.
+    data_collator = FlaxDataCollatorForT5MLM(
+        tokenizer=tokenizer,
+        noise_density=data_args.mlm_probability,
+        mean_noise_span_length=data_args.mean_noise_span_length,
+        input_length=max_seq_length,
+        target_length=targets_length,
+        pad_token_id=model.config.pad_token_id,
+        decoder_start_token_id=model.config.decoder_start_token_id,
+    )
+
+    # 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()
+
+    num_train_steps = data_args.num_train_steps
+
+    # Create learning rate schedule
+    warmup_fn = optax.linear_schedule(
+        init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps
+    )
+    decay_fn = optax.linear_schedule(
+        init_value=training_args.learning_rate,
+        end_value=0,
+        transition_steps=num_train_steps - training_args.warmup_steps,
+    )
+    linear_decay_lr_schedule_fn = optax.join_schedules(
+        schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]
+    )
+
+    # 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.
+    def decay_mask_fn(params):
+        flat_params = traverse_util.flatten_dict(params)
+        flat_mask = {
+            path: (path[-1] != "bias" and path[-2:] not in [("layer_norm", "scale"), ("final_layer_norm", "scale")])
+            for path in flat_params
+        }
+        return traverse_util.unflatten_dict(flat_mask)
+
+    # create adam optimizer
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
+
+    # Setup train state
+    state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer)
+
+    # Define gradient update step fn
+    def train_step(state, batch, dropout_rng):
+        dropout_rng, new_dropout_rng = jax.random.split(dropout_rng)
+
+        def loss_fn(params):
+            labels = batch.pop("labels")
+
+            logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0]
+
+            # compute loss
+            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])).mean()
+
+            return loss
+
+        grad_fn = jax.value_and_grad(loss_fn)
+        loss, grad = grad_fn(state.params)
+        grad = jax.lax.pmean(grad, "batch")
+        new_state = state.apply_gradients(grads=grad)
+
+        metrics = jax.lax.pmean(
+            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch"
+        )
+
+        return new_state, metrics, new_dropout_rng
+
+    # Create parallel version of the train step
+    p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0,))
+
+    # Define eval fn
+    def eval_step(params, batch):
+        labels = batch.pop("labels")
+
+        logits = model(**batch, params=params, train=False)[0]
+
+        # compute loss
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1]))
+
+        # compute accuracy
+        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels)
+
+        # summarize metrics
+        metrics = {"loss": loss.mean(), "accuracy": accuracy.mean()}
+        metrics = jax.lax.pmean(metrics, axis_name="batch")
+
+        return metrics
+
+    p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0,))
+
+    # Replicate the train state on each device
+    state = jax_utils.replicate(state)
+
+    train_time = 0
+    train_start = time.time()
+    train_metrics = []
+    eval_metrics = []
+
+    training_iter = iter(tokenized_datasets)
+
+    eval_samples = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, expanded_inputs_length)
+
+    steps = tqdm(range(num_train_steps), desc="Training...", position=0)
+    for step in range(num_train_steps):
+        # ======================== Training ================================
+        try:
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, expanded_inputs_length)
+        except StopIteration:
+            # Once the end of the dataset stream is reached, the training iterator
+            # is reinitialized and reshuffled and a new eval dataset is randomely chosen.
+            shuffle_seed += 1
+            tokenized_datasets.set_epoch(shuffle_seed)
+
+            training_iter = iter(tokenized_datasets)
+
+            eval_dataset = advance_iter_and_group_samples(training_iter, data_args.num_eval_samples, expanded_inputs_length)
+            samples = advance_iter_and_group_samples(training_iter, train_batch_size, expanded_inputs_length)
+
+        # Model forward
+        model_inputs = data_collator(samples)
+        model_inputs = shard(model_inputs.data)
+        state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
+        train_metrics.append(train_metric)
+
+        if step % training_args.logging_steps == 0 and step > 0:
+            train_metric = jax_utils.unreplicate(train_metric)
+            steps.write(
+                f"Step... ({step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})"
+            )
+            train_time += time.time() - train_start
+            if has_tensorboard and jax.process_index() == 0:
+                write_train_metric(summary_writer, train_metrics, train_time, step)
+            train_metrics = []
+        # ======================== Evaluating ==============================
+        if step % training_args.eval_steps == 0 and step > 0:
+            eval_samples_idx = jnp.arange(data_args.num_eval_samples)
+            eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+
+            for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=1)):
+                # process input samples
+                batch_eval_samples = {k: [v[idx] for idx in batch_idx] for k, v in eval_samples.items()}
+                model_inputs = data_collator(batch_eval_samples)
+
+                # Model forward
+                model_inputs = shard(model_inputs.data)
+                metrics = p_eval_step(state.params, model_inputs)
+                eval_metrics.append(metrics)
+
+            # normalize eval metrics
+            eval_metrics = get_metrics(eval_metrics)
+            eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+
+            # Update progress bar
+            steps.desc = f"Step... ({step + 1}/{num_train_steps} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+            if has_tensorboard and jax.process_index() == 0:
+                write_eval_metric(summary_writer, eval_metrics, step)
+            eval_metrics = []
+
+            if step % training_args.save_steps == 0 and step > 0:
+            # save checkpoint after each save_steps 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 step {step+1}",
+                    )
+                    #tokenizer.save_pretrained(
+                    #    training_args.output_dir
+                    #)
+
+        # update tqdm bar
+        steps.update(1)