run_lfom_distillation_flax.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace Team All rights reserved.
# Copyright 2022 Vladislav Lialin
#
# 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.
"""
Distilling T5 model using learning from other's mistakes (https://arxiv.org/abs/2012.01300) approach.

Here is the full list of checkpoints on the hub that can be pretrained by this script:
https://huggingface.co/models?filter=t5

Based on https://github.com/huggingface/transformers/blob/0f69b924fbda6a442d721b10ece38ccfc6b67275/examples/flax/language-modeling/run_mlm_flax.py
"""
import json
import logging
import os
import sys
import time
from functools import partial
from dataclasses import asdict, dataclass, field

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

import numpy as np
import datasets as hf_datasets
from datasets import load_dataset
from tqdm import tqdm

import flax
import jax
import jax.numpy as jnp
import optax
import flax.serialization
from flax import jax_utils, traverse_util
from flax.training import train_state
from flax.training.common_utils import get_metrics, onehot, shard
from huggingface_hub import Repository
import transformers
from transformers import (
    CONFIG_MAPPING,
    FLAX_MODEL_FOR_MASKED_LM_MAPPING,
    AutoTokenizer,
    FlaxT5ForConditionalGeneration,
    HfArgumentParser,
    T5Config,
    set_seed,
)
from transformers.file_utils import get_full_repo_name
import wandb

from lfom_distillation.utils import (
    compute_input_and_target_lengths,
    FlaxDataCollatorForT5MLM,
    generate_batch_splits,
)


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


logging.basicConfig(
    format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    level=logging.INFO,
    stream=sys.stdout,
)
logger = logging.getLogger(__name__)
hf_datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_warning()


@dataclass
class TrainingArguments:
    output_dir: str = field(
        metadata={"help": "The output directory where the model predictions and checkpoints will be written."},
    )
    overwrite_output_dir: bool = field(
        default=False,
        metadata={
            "help": (
                "Overwrite the content of the output directory. "
                "Use this to continue training if output_dir points to a checkpoint directory."
            )
        },
    )
    do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
    do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
    per_device_train_batch_size: int = field(
        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for training."}
    )
    per_device_eval_batch_size: int = field(
        default=8, metadata={"help": "Batch size per GPU/TPU core/CPU for evaluation."}
    )
    temperature: float = field(default=1.0, metadata={"help": "Distillation temperature used in softmax."})
    learning_rate: float = field(default=5e-5, metadata={"help": "The initial learning rate for AdamW."})
    gradient_accumulation_steps: int = field(default=1, metadata={"help": "Number of grad acc steps."})
    weight_decay: float = field(default=0.0, metadata={"help": "Weight decay for AdamW if we apply some."})
    adam_beta1: float = field(default=0.9, metadata={"help": "Beta1 for AdamW optimizer"})
    adam_beta2: float = field(default=0.999, metadata={"help": "Beta2 for AdamW optimizer"})
    adam_epsilon: float = field(default=1e-8, metadata={"help": "Epsilon for AdamW optimizer."})
    adafactor: bool = field(default=False, metadata={"help": "Whether or not to replace AdamW by Adafactor."})
    num_train_epochs: float = field(default=3.0, metadata={"help": "Total number of training epochs to perform."})
    num_train_steps: int = field(default=None, metadata={"help": "Overrides the number of steps defined by `num_train_epochs`"})
    skip_train_steps: int = field(default=0, metadata={"help": "Skip this many training steps. Can be used to restart the training after a failure."})
    warmup_steps: int = field(default=0, metadata={"help": "Linear warmup over warmup_steps."})
    logging_steps: int = field(default=500, metadata={"help": "Log every X updates steps."})
    save_steps: int = field(default=500, metadata={"help": "Save checkpoint every X updates steps."})
    eval_steps: int = field(default=None, metadata={"help": "Run an evaluation every X steps."})
    seed: int = field(default=42, metadata={"help": "Random seed that will be set at the beginning of training."})
    push_to_hub: bool = field(
        default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
    )
    hub_model_id: str = field(
        default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
    )
    hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})

    def __post_init__(self):
        if self.output_dir is not None:
            self.output_dir = os.path.expanduser(self.output_dir)

    def to_dict(self):
        """
        Serializes this instance while replace `Enum` by their values (for JSON serialization support). It obfuscates
        the token values by removing their value.
        """
        d = asdict(self)
        for k, v in d.items():
            if isinstance(v, Enum):
                d[k] = v.value
            if isinstance(v, list) and len(v) > 0 and isinstance(v[0], Enum):
                d[k] = [x.value for x in v]
            if k.endswith("_token"):
                d[k] = f"<{k.upper()}>"
        return d


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

    teacher_model_name_or_path: str = field(
        metadata={"help": "The model checkpoint for teacher model weights."}
    )

    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."
        },
    )

    weak_model_name_or_path: Optional[str] = field(
        default=None,
        metadata={"help": "The model checkpoint for weak model weights that is used for learning-from-others' mistakes."},
    )

    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"
        },
    )
    dataset_fraction: Optional[float] = field(
        default=1.0, metadata={"help": "The fraction of the dataset to use for training (if < 1.0)."}
    )
    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"},
    )

    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", "txt"], "`train_file` should be a csv, a json or a txt file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."


class DistillationTrainState(train_state.TrainState):
    teacher_params: flax.core.FrozenDict = None
    weak_params: flax.core.FrozenDict = None

    @classmethod
    def create(cls, *, apply_fn, params, tx, teacher_params, weak_params=None):
        """Creates a new instance with `step=0` and initialized `opt_state`."""
        opt_state = tx.init(params)
        return cls(
            step=0,
            apply_fn=apply_fn,
            params=params,
            tx=tx,
            opt_state=opt_state,
            teacher_params=teacher_params,
            weak_params=weak_params,
        )


def write_train_metric(train_metrics, train_time, step):
    train_metrics = get_metrics(train_metrics)

    # only keep the last value for each metric,
    # this function should be called often enough to not worry about noisy values
    train_metrics = {
        f"train_{metric_name}": values[-1] for metric_name, values in train_metrics.items()
    }
    wandb.log(train_metrics, step=step)


def write_eval_metric(eval_metrics, step):
    for metric_name, value in eval_metrics.items():
        wandb.log({f"eval_{metric_name}": value}, step=step)


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."
        )

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
        level="NOTSET",
        datefmt="[%X]",
    )
    # ignore urllib3.connectionpool
    logging.getLogger("urllib3.connectionpool").setLevel(logging.WARNING)

    # 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)

    # Handle the repository creation
    if training_args.push_to_hub:
        if training_args.hub_model_id is None:
            repo_name = get_full_repo_name(
                Path(training_args.output_dir).absolute().name, token=training_args.hub_token
            )
        else:
            repo_name = training_args.hub_model_id
        repo = Repository(training_args.output_dir, clone_from=repo_name)

    # 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)

        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)

        if "validation" not in datasets.keys():
            datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
            )
            datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
            )

    # Sample data if needed (used for debugging)
    if data_args.dataset_fraction < 1.0:
        for k, v in datasets.items():
            indices = np.random.choice(len(v), int(len(v) * data_args.dataset_fraction), replace=False)
            datasets[k] = v.select(indices)

    # 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 = 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 model_args.config_name:
        config = T5Config.from_pretrained(
            model_args.config_name, cache_dir=model_args.cache_dir,
        )
    elif model_args.model_name_or_path:
        config = T5Config.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.")

    # 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)

    # 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[text_column_name], return_attention_mask=False)

    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,
    )

    # 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,
    )

    # Main data processing function that will concatenate all texts from our dataset and generate chunks of expanded_inputs_length.
    def group_texts(examples):
        # Concatenate all texts.
        concatenated_examples = {k: list(chain(*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 >= expanded_inputs_length:
            total_length = (total_length // expanded_inputs_length) * expanded_inputs_length
        # Split by chunks of max_len.
        result = {
            k: [t[i : i + expanded_inputs_length] for i in range(0, total_length, expanded_inputs_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
    tokenized_datasets = tokenized_datasets.map(
        group_texts,
        batched=True,
        num_proc=data_args.preprocessing_num_workers,
        load_from_cache_file=not data_args.overwrite_cache,
    )

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

    teacher_model = FlaxT5ForConditionalGeneration.from_pretrained(
        model_args.teacher_model_name_or_path, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
    )

    if model_args.model_name_or_path:
        student_model = FlaxT5ForConditionalGeneration.from_pretrained(
            model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )
    else:
        student_model = FlaxT5ForConditionalGeneration(config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype))

    weak_model = None
    if model_args.weak_model_name_or_path:
        weak_model = FlaxT5ForConditionalGeneration.from_pretrained(
            model_args.weak_model_name_or_path, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
        )

    # Enable wandb only on the master node
    if jax.process_index() == 0:
        # note that if same key exists in multiple args, it will be overwritten by the last one
        hparams = {
            **vars(training_args),
            **vars(model_args),
            **vars(data_args),
            **student_model.config.to_dict(),
        }
        wandb.init(project="t5-lm-lfom-distillation", config=hparams)

    # 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=student_model.config.pad_token_id,
        decoder_start_token_id=student_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 = len(tokenized_datasets["train"]) // train_batch_size * num_epochs

    if training_args.num_train_steps is not None:
        num_train_steps = training_args.num_train_steps
        num_batches_in_epoch = len(tokenized_datasets["train"]) // train_batch_size
        num_epochs = num_train_steps // num_batches_in_epoch + int(num_train_steps % num_batches_in_epoch > 0)

    logger.info(f"Dataset size is {len(tokenized_datasets['train'])} train examples.")
    logger.info(f"Given batch size {train_batch_size}, number of batches per epoch is {len(tokenized_datasets['train']) // train_batch_size}.")
    logger.info(f"Training will run for {num_train_steps} steps ({num_epochs} 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.
    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,
        )

    if training_args.gradient_accumulation_steps:
        optimizer = optax.MultiSteps(optimizer, every_k_schedule=training_args.gradient_accumulation_steps)

    # Setup train state
    def apply_distillation(student_params, teacher_params, student_model_fn, teacher_model_fn, temperature, weak_params=None, weak_model_fn=None, **batch):
        student_logits = student_model_fn(params=student_params, **batch)[0]
        teacher_logits = teacher_model_fn(params=teacher_params, **batch)[0]

        if weak_model_fn:
            assert weak_params is not None
            weak_logits = weak_model_fn(params=weak_params, **batch)[0]
            weak_logits = jax.lax.stop_gradient(weak_logits)
            student_logits = student_logits + weak_logits

        student_logits = student_logits / temperature
        teacher_logits = teacher_logits / temperature
        teacher_probs = jax.nn.softmax(teacher_logits)
        teacher_probs = jax.lax.stop_gradient(teacher_probs)

        # compute loss
        loss = optax.softmax_cross_entropy(student_logits, teacher_probs).mean()
        loss *= temperature ** 2

        return loss, student_logits

    apply = partial(
        apply_distillation,
        student_model_fn=student_model.__call__,
        teacher_model_fn=teacher_model.__call__,
        weak_model_fn=weak_model.__call__ if weak_model else None,
        temperature=0.5,
    )

    state = DistillationTrainState.create(
        apply_fn=apply,
        params=student_model.params,
        teacher_params=teacher_model.params,
        weak_params=weak_model.params if weak_model else None,
        tx=optimizer,
    )

    # load state if exists
    if training_args.skip_train_steps > 0:
        logger.info("Trying to resume training, setting state.steps to skip_train_step and loading optimizer state from optax checkpoint.")
        state = state.replace(step=training_args.skip_train_steps)

        opt_path = os.path.join(training_args.output_dir, "opt_state.msgpack")
        if os.path.exists(opt_path):
            with open(opt_path, "rb") as f:
                opt_state = flax.serialization.from_bytes(optax.MultiStepsState, f.read())
                opt_state = optax.MultiStepsState(
                    mini_step=opt_state["mini_step"],
                    gradient_step=opt_state["gradient_step"],
                    inner_opt_state=opt_state["inner_opt_state"],
                    acc_grads=opt_state["acc_grads"],
                )

            state = state.replace(opt_state=opt_state)

        else:
            logger.warning("Optimizer state not found, expect loss to explode.")

    # 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(student_params, teacher_params):
            labels = batch.pop("labels")  # we do not use labels, because distillation uses teacher logits

            # TODO: double check that it works, because batch is not a part of this function
            loss, _ = state.apply_fn(
                **batch,
                student_params=student_params,
                teacher_params=teacher_params,
                weak_params=state.weak_params,
                dropout_rng=dropout_rng,
                train=True,
            )

            return loss

        grad_fn = jax.value_and_grad(loss_fn, argnums=0)  # only differentiate with respect to the first argument (student_params)
        loss, grad = grad_fn(state.params, state.teacher_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 = student_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
    epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0)

    for epoch in epochs:
        # ======================== Training ================================
        train_start = time.time()
        train_metrics = []

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

        # Generate an epoch by shuffling sampling indices from the train dataset
        num_train_samples = len(tokenized_datasets["train"])
        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)

        # 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)):
            cur_step = epoch * (num_train_samples // train_batch_size) + step

            if cur_step < training_args.skip_train_steps:
                # note that we have updated state.step above
                continue

            samples = [tokenized_datasets["train"][int(idx)] for idx in batch_idx]
            model_inputs = data_collator(samples)

            # 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)

            if cur_step % training_args.logging_steps == 0 and cur_step > 0:
                # Save metrics
                train_metric = jax_utils.unreplicate(train_metric)
                train_time += time.time() - train_start
                if jax.process_index() == 0:
                    write_train_metric(train_metrics, train_time, cur_step)

                train_metrics = []

            if cur_step % training_args.eval_steps == 0 and cur_step > 0:
                # ======================== Evaluating ==============================
                num_eval_samples = len(tokenized_datasets["validation"])
                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 = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
                    model_inputs = data_collator(samples)

                    # Model forward
                    model_inputs = shard(model_inputs.data)
                    metrics = p_eval_step(state.params, model_inputs)
                    eval_metrics.append(metrics)

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

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

            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))
                    student_model.save_pretrained(training_args.output_dir, params=params)
                    tokenizer.save_pretrained(training_args.output_dir)

                    opt_state = jax.device_get(jax.tree_map(lambda x: x[0], state.opt_state))
                    opt_path = os.path.join(training_args.output_dir, "opt_state.msgpack")
                    with open(opt_path, "wb") as f:
                        opt_bytes = flax.serialization.to_bytes(opt_state)
                        f.write(opt_bytes)

                    if training_args.push_to_hub:
                        repo.push_to_hub(commit_message=f"Saving weights and logs of step {cur_step}", blocking=False)

            if training_args.num_train_steps is not None and cur_step > training_args.num_train_steps:
                logger.info("Reached maximum number of training steps")
                break

    # Eval after training
    if training_args.do_eval:
        logger.info(f"Running final evaluation. The model is saved to {training_args.output_dir}")
        num_eval_samples = len(tokenized_datasets["validation"])
        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 = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
            model_inputs = data_collator(samples)

            # Model forward
            model_inputs = shard(model_inputs.data)
            metrics = p_eval_step(state.params, model_inputs)
            eval_metrics.append(metrics)

        # get eval metrics
        eval_metrics = get_metrics(eval_metrics)
        eval_metrics = jax.tree_map(lambda metric: jnp.mean(metric).item(), eval_metrics)

        if jax.process_index() == 0:
            eval_metrics = {f"eval_{metric_name}": value for metric_name, value in eval_metrics.items()}
            path = os.path.join(training_args.output_dir, "eval_results.json")
            with open(path, "w") as f:
                json.dump(eval_metrics, f, indent=4, sort_keys=True)


if __name__ == "__main__":
    main()