run_mlm_flax.py

#!/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 masked language modeling (BERT, ALBERT, RoBERTa...) with whole word masking on a
text file or a dataset.

Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=masked-lm
"""
import shutil
import logging
import os
import sys
import time
from dataclasses import dataclass, field
from ast import Str

# You can also adapt this script on your own masked language modeling task. Pointers for this are left as comments.
from pathlib import Path
from typing import Dict, List, Optional, Tuple

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,
    AutoConfig,
    AutoTokenizer,
    FlaxAutoModelForMaskedLM,
    HfArgumentParser,
    PreTrainedTokenizerBase,
    TensorType,
    TrainingArguments,
    is_tensorboard_available,
    set_seed,
)
from transformers.testing_utils import CaptureLogger
from flax.serialization import to_bytes, from_bytes
from importlib.util import find_spec
from flax.training import checkpoints
from flax.jax_utils import unreplicate
from flax.training.checkpoints import save_checkpoint, restore_checkpoint
import json

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]`."
        },
    )

    


@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)."}
    )
    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)."},
    )
    train_ref_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input train ref data file for whole word masking in Chinese."},
    )
    validation_ref_file: Optional[str] = field(
        default=None,
        metadata={"help": "An optional input validation ref data file for whole word masking in Chinese."},
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    validation_split_percentage: Optional[int] = field(
        default=5,
        metadata={
            "help": "The percentage of the train set used as validation set in case there's no validation split"
        },
    )
    max_seq_length: Optional[int] = field(
        default=None,
        metadata={
            "help": "The maximum total input sequence length after tokenization. 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 masked language modeling loss"}
    )
    pad_to_max_length: bool = field(
        default=False,
        metadata={
            "help": "Whether to pad all samples to `max_seq_length`. "
            "If False, will pad the samples dynamically when batching to the maximum length in the batch."
        },
    )
    line_by_line: bool = field(
        default=False,
        metadata={"help": "Whether distinct lines of text in the dataset are to be handled as distinct sequences."},
    )


@flax.struct.dataclass
class FlaxDataCollatorForLanguageModeling:
    """
    Data collator used for language modeling. Inputs are dynamically padded to the maximum length of a batch if they
    are not all of the same length.

    Args:
        tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`):
            The tokenizer used for encoding the data.
        mlm_probability (:obj:`float`, `optional`, defaults to 0.15):
            The probability with which to (randomly) mask tokens in the input.

    .. note::

        For best performance, this data collator should be used with a dataset having items that are dictionaries or
        BatchEncoding, with the :obj:`"special_tokens_mask"` key, as returned by a
        :class:`~transformers.PreTrainedTokenizer` or a :class:`~transformers.PreTrainedTokenizerFast` with the
        argument :obj:`return_special_tokens_mask=True`.
    """

    tokenizer: PreTrainedTokenizerBase
    mlm_probability: float = 0.15

    def __post_init__(self):
        if self.tokenizer.mask_token is None:
            raise ValueError(
                "This tokenizer does not have a mask token which is necessary for masked language modeling. "
                "You should pass `mlm=False` to train on causal language modeling instead."
            )

    def __call__(self, examples: List[Dict[str, np.ndarray]], pad_to_multiple_of: int) -> Dict[str, np.ndarray]:
        # Handle dict or lists with proper padding and conversion to tensor.
        batch = self.tokenizer.pad(examples, pad_to_multiple_of=pad_to_multiple_of, return_tensors=TensorType.NUMPY)

        # If special token mask has been preprocessed, pop it from the dict.
        special_tokens_mask = batch.pop("special_tokens_mask", None)

        batch["input_ids"], batch["labels"] = self.mask_tokens(
            batch["input_ids"], special_tokens_mask=special_tokens_mask
        )
        return batch

    def mask_tokens(
        self, inputs: np.ndarray, special_tokens_mask: Optional[np.ndarray]
    ) -> Tuple[jnp.ndarray, jnp.ndarray]:
        """
        Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.
        """
        labels = inputs.copy()
        # We sample a few tokens in each sequence for MLM training (with probability `self.mlm_probability`)
        probability_matrix = np.full(labels.shape, self.mlm_probability)
        special_tokens_mask = special_tokens_mask.astype("bool")

        probability_matrix[special_tokens_mask] = 0.0
        masked_indices = np.random.binomial(1, probability_matrix).astype("bool")
        labels[~masked_indices] = -100  # We only compute loss on masked tokens

        # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])
        indices_replaced = np.random.binomial(1, np.full(labels.shape, 0.8)).astype("bool") & masked_indices
        inputs[indices_replaced] = self.tokenizer.convert_tokens_to_ids(self.tokenizer.mask_token)

        # 10% of the time, we replace masked input tokens with random word
        indices_random = np.random.binomial(1, np.full(labels.shape, 0.5)).astype("bool")
        indices_random &= masked_indices & ~indices_replaced

        random_words = np.random.randint(self.tokenizer.vocab_size, size=labels.shape, dtype="i4")
        inputs[indices_random] = random_words[indices_random]

        # The rest of the time (10% of the time) we keep the masked input tokens unchanged
        return inputs, labels


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 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)

def mb_item(x):
    return x.item() if hasattr(x, "item") else x

#checkpoint functions





def rotate_checkpoints(ckpt_dir: str, save_total_limit: int):
    "Removes older checkpoints so that `save_total_limit` checkpoints are kept"
    # TODO: what to remove is decided using step number only, we might want to improve that
    ckpts = [str(x) for x in Path(ckpt_dir).glob("ckpt-*")]
    # sort checkpoints by step
    ckpts_sorted = sorted(ckpts, key=lambda x: int(x.split('-')[-1]))
    ckpts_to_delete = ckpts_sorted[:-save_total_limit]
    for ckpt in ckpts_to_delete:
        logger.info(f"Deleting older checkpoint [{ckpt}] due to save_total_limit ({save_total_limit})")
        shutil.rmtree(ckpt)


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


    
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="NOTSET",
        datefmt="[%X]",
    )

    # Log on each process the small summary:
    logger = logging.getLogger(__name__)

    # 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).
    #
    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    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)

        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )
    else:
        #data_files = {}
        #if data_args.train_file is not None:
        #    data_files["train"] = data_args.train_file
        #if data_args.validation_file is not None:
        #    data_files["validation"] = data_args.validation_file
        #extension = data_args.train_file.split(".")[-1]
        #if extension == "txt":
        #    extension = "text"
        #datasets = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir)

        #data_dir = "/home/yeb"
        # data_dir = "/home/yeb/Developer/data"
        data_files = []
        def train_val_files():
            import glob
            import random
            SEED = 42
            def add_jsonlines_dir(path):
                global data_files
                data_files += glob.glob(f"{path}/*.gz")
                
            add_jsonlines_dir("/home/dat/subset_c4_cleannl")   
            add_jsonlines_dir("/data/oscar_nl_cleaned")
            add_jsonlines_dir("/data/nrc_cleaned_idtextfmt")
            add_jsonlines_dir("/data/nu_cleaned_idtextfmt")
            random.Random(SEED).shuffle(data_files)
            total = len(data_files)
            val_size = int(0.05 * total)
            train_size = total - val_size
            print(f"95%: {train_size}")
            train = data_files[:train_size]
            val = data_files[train_size:]
            print(f"Got {len(train)} training files and {len(val)} validation files")
            assert list(set(train) & set(val)) == [], "Train overlaps with test"
            return train, val
        train, val = train_val_files()
        datasets = load_dataset('json', data_files={'train': train, 'validation': val})
        datasets["train"] = datasets["train"].select(range(int(0.8*len(datasets["train"]))))
        datasets["validation"] = datasets["validation"].select(range(int(0.8*len(datasets["validation"]))))
        #datasets["train"] = datasets["train"].select(range(10000))
        #datasets["validation"] = datasets["validation"].select(range(10000))





    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."
        )

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    if training_args.do_train:
        column_names = datasets["train"].column_names
    else:
        column_names = datasets["validation"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]

    max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length)


    if data_args.line_by_line:
        # When using line_by_line, we just tokenize each nonempty line.
        padding = "max_length" if data_args.pad_to_max_length else False

        def tokenize_function(examples):
            # Remove empty lines
            examples = [line for line in examples if len(line) > 0 and not line.isspace()]
            return tokenizer(
                examples,
                return_special_tokens_mask=True,
                padding=padding,
                truncation=True,
                max_length=max_seq_length,
            )

        tokenized_datasets = datasets.map(
            tokenize_function,
            input_columns=[text_column_name],
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
        )

    else:
        # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts.
        # We use `return_special_tokens_mask=True` because DataCollatorForLanguageModeling (see below) is more
        # efficient when it receives the `special_tokens_mask`.
        def tokenize_function(examples):
            return tokenizer(examples[text_column_name], return_special_tokens_mask=True)

        tokenized_datasets = datasets.map(
            tokenize_function,
            batched=True,
            num_proc=data_args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not data_args.overwrite_cache,
        )

        # Main data processing function that will concatenate all texts from our dataset and generate chunks of
        # max_seq_length.
        def group_texts(examples):
            # Concatenate all texts.
            concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
            total_length = len(concatenated_examples[list(examples.keys())[0]])
            # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
            # customize this part to your needs.
            if total_length >= max_seq_length:
                total_length = (total_length // max_seq_length) * max_seq_length
            # Split by chunks of max_len.
            result = {
                k: [t[i : i + max_seq_length] for i in range(0, total_length, max_seq_length)]
                for k, t in concatenated_examples.items()
            }
            return result

        # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a
        # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value
        # might be slower to preprocess.
        #
        # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
        # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
        lm_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            batch_size=100,
            num_proc=data_args.preprocessing_num_workers,
            load_from_cache_file=not data_args.overwrite_cache,
        )
        train_dataset = lm_datasets["train"]
        eval_dataset = lm_datasets["validation"]
        



    # 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."
        )
    # enable wandb tracking
    has_wandb = find_spec("wandb") is not None 
    if jax.process_index() == 0 and has_wandb and ("wandb" in training_args.report_to):
        try:
            import wandb
            wandb.init(
                entity="wandb", 
                project="hf-flax-pino-roberta",
                sync_tensorboard=True
            )
            wandb.config.update(training_args)
            wandb.config.update(model_args)
            wandb.config.update(data_args)
        except ImportError as e:
            print(e)
            has_wandb = False
    
    # Data collator
    # This one will take care of randomly masking the tokens.
    data_collator = FlaxDataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=data_args.mlm_probability)

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

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

    # 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 = len(train_dataset) // train_batch_size * num_epochs

    # 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.
    # Note that this mask is specifically adapted for FlaxBERT-like models.
    # For other models, one should correct the layer norm parameter naming
    # accordingly.
    def decay_mask_fn(params):
        flat_params = traverse_util.flatten_dict(params)
        flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "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,
            eps=training_args.adam_epsilon,
            weight_decay=training_args.weight_decay,
            mask=decay_mask_fn,
        )

    #if training_args.gradient_accumulation_steps > 1:
    #    optimizer = optax.MultiSteps(optimizer, training_args.gradient_accumulation_steps)
    #grad_accum_steps = training_args.gradient_accumulation_steps

    # Setup train state
    
    
    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer,grad_accum=jax.tree_map(jnp.zeros_like, model.params))
        
    if training_args.resume_from_checkpoint:
        state = restore_checkpoint(training_args.resume_from_checkpoint, state)
        resume_step = mb_item(state.step.item())
    else:
        resume_step = 0
    

    # 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, ignore padded input tokens
            label_mask = jnp.where(labels > 0, 1.0, 0.0)
            loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask

            # take average
            loss = loss.sum() / label_mask.sum()

            return loss / training_args.gradient_accumulation_steps

        grad_fn = jax.value_and_grad(loss_fn)
        loss, grads = grad_fn(state.params)
        grad_accum = jax.tree_multimap(lambda x, y: x + y, grads, state.grad_accum)
        
        def update_fn():
            grads = jax.tree_map(lambda x: x / training_args.gradient_accumulation_steps, grad_accum)
            grads = jax.lax.pmean(grad_accum, "batch")
            new_state = state.apply_gradients(grads=grads,grad_accum=jax.tree_map(jnp.zeros_like, grads))
            return new_state

        new_state = jax.lax.cond(
            state.step % training_args.gradient_accumulation_steps == 0,
            lambda _: update_fn(),
            lambda _: state.replace(grad_accum=grad_accum, step=state.step + 1),
            None,
        )
        
        metrics = jax.lax.pmean(
            {"loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step)}, axis_name="batch" #
        )

        #return new_state.replace(new_dropout_rng=new_dropout_rng), metrics
        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, ignore padded input tokens
        label_mask = jnp.where(labels > 0, 1.0, 0.0)
        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) * label_mask

        # compute accuracy
        accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) * label_mask

        # summarize metrics
        metrics = {"loss": loss.sum(), "accuracy": accuracy.sum(), "normalizer": label_mask.sum()}
        metrics = jax.lax.psum(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
    epochs = tqdm(range(num_epochs), desc=f"Epoch ... (1/{num_epochs})", position=0)
    for epoch in epochs:
        # ======================== Training ================================
        train_start = time.time()
        train_metrics = []

        # Create sampling rng
        rng, input_rng = jax.random.split(rng)
        steps_per_epoch = len(train_dataset) // train_batch_size

        # Generate an epoch by shuffling sampling indices from the train dataset
        num_train_samples = len(train_dataset)
        train_samples_idx = jax.random.permutation(input_rng, jnp.arange(num_train_samples))
        train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size) #// grad_accum_steps

        # Gather the indexes for creating the batch and do a training step
        for step, batch_idx in enumerate(tqdm(train_batch_idx, desc="Training...", position=1,initial=resume_step)): #grad_accum
            samples = [train_dataset[int(idx)] for idx in batch_idx]
            model_inputs = data_collator(samples, pad_to_multiple_of=16)
            

            # Model forward
            model_inputs = shard(model_inputs.data)
            state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
            train_metrics.append(train_metric)

            cur_step = epoch * (num_train_samples // train_batch_size) + step
            if cur_step < resume_step:
                continue

            if (cur_step % training_args.logging_steps) == 0 and cur_step > 0: # * grad_accum_steps
                # Save metrics
                train_metric = jax_utils.unreplicate(train_metric)
                train_time += time.time() - train_start
                if has_tensorboard and jax.process_index() == 0:
                    write_train_metric(summary_writer, train_metrics, train_time, cur_step)
                if has_wandb and jax.process_index() == 0 and ("wandb" in training_args.report_to):
                    # TODO: add accumulation of metrics
                    _metrics = {k if k=="learning_rate" else f"train_{k}":mb_item(v.mean()) for k, v in train_metric.items()}
                    wandb.log({"training_step":cur_step, **_metrics}, commit=True)
                
                epochs.write(
                    f"Step... ({cur_step} | Loss: {train_metric['loss']}, Learning Rate: {train_metric['learning_rate']})"
                )

                train_metrics = []

            if cur_step % (training_args.eval_steps) == 0 and cur_step > 0: #* grad_accum_steps
                # ======================== Evaluating ==============================
                num_eval_samples = len(eval_dataset)
                eval_samples_idx = jnp.arange(num_eval_samples)
                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)

                eval_metrics = []
                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
                    samples = [eval_dataset[int(idx)] for idx in batch_idx]
                    model_inputs = data_collator(samples, pad_to_multiple_of=16)

                    # 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)
                eval_normalizer = eval_metrics.pop("normalizer")
                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)

                # Update progress bar
                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"

                # Save metrics
                if has_tensorboard and jax.process_index() == 0:
                    write_eval_metric(summary_writer, eval_metrics, cur_step)

                if has_wandb and jax.process_index() == 0 and ("wandb" in training_args.report_to):
                    _metrics = {f"eval_{k}":mb_item(v) for k, v in eval_metrics.items()}
                    wandb.log({"eval_step":cur_step, **_metrics})

            if (cur_step % training_args.save_steps == 0 ) and cur_step > 0: #
                # 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 step {cur_step}",
                    )
                    save_checkpoint(training_args.output_dir, jax_utils.unreplicate(state), cur_step, keep=training_args.save_total_limit, overwrite=True)
                    if training_args.save_total_limit is not None:
                        rotate_checkpoints(training_args.output_dir, training_args.save_total_limit)

    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 {cur_step}",
            )