docs/transformers/examples/pytorch/language-modeling/run_fim.py

#!/usr/bin/env python
# Copyright 2024 The HuggingFace Inc. 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 causal language modeling using
Fill-in-the middle (FIM) objective 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=text-generation
"""
# You should adapt this script on your own causal language modeling task. Pointers for this are left as comments.

import logging
import math
import os
import sys
from dataclasses import dataclass, field
from itertools import chain
from typing import Optional

import datasets
import evaluate
import numpy as np
import torch
from datasets import load_dataset

import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_FOR_CAUSAL_LM_MAPPING,
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    default_data_collator,
    is_torch_xla_available,
    set_seed,
)
from transformers.integrations import is_deepspeed_zero3_enabled
from transformers.testing_utils import CaptureLogger
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version


# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.52.0.dev0")

require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")

logger = logging.getLogger(__name__)


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


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

    model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The model checkpoint for weights initialization. Don't set if you want to train a model from scratch."
            )
        },
    )
    model_type: Optional[str] = field(
        default=None,
        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
    )
    config_overrides: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override some existing default config settings when a model is trained from scratch. Example: "
                "n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
            )
        },
    )
    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 huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=True,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    token: str = field(
        default=None,
        metadata={
            "help": (
                "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
                "generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
            )
        },
    )
    trust_remote_code: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to trust the execution of code from datasets/models defined on the Hub."
                " This option should only be set to `True` for repositories you trust and in which you have read the"
                " code, as it will execute code present on the Hub on your local machine."
            )
        },
    )
    torch_dtype: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
                "dtype will be automatically derived from the model's weights."
            ),
            "choices": ["auto", "bfloat16", "float16", "float32"],
        },
    )
    low_cpu_mem_usage: bool = field(
        default=False,
        metadata={
            "help": (
                "It is an option to create the model as an empty shell, then only materialize its parameters when the pretrained weights are loaded. "
                "set True will benefit LLM loading time and RAM consumption."
            )
        },
    )
    pad_to_multiple_of: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to pad the embedding layer to a multiple depending on the device. ",
                "For NVIDIA GPUs, this will be a multiple of 8, for TPUs a multiple of 128.",
            )
        },
    )
    attn_implementation: Optional[str] = field(
        default="sdpa", metadata={"help": ("The attention implementation to use. ")}
    )

    def __post_init__(self):
        if self.config_overrides is not None and (self.config_name is not None or self.model_name_or_path is not None):
            raise ValueError(
                "--config_overrides can't be used in combination with --config_name or --model_name_or_path"
            )


@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)."},
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
        },
    )
    streaming: bool = field(default=False, metadata={"help": "Enable streaming mode"})
    block_size: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "Optional input sequence length after tokenization. "
                "The training dataset will be truncated in block of this size for training. "
                "Default to the model max input length for single sentence inputs (take into account special tokens)."
            )
        },
    )
    fim_rate: Optional[float] = field(
        default=0.5,
        metadata={
            "help": (
                "Optional probability with which the FIM transformation is applied to the example. "
                "Default is 0.5. A rate of 1.0 means every example will undergo FIM transformation, "
                "while a rate of 0.0 means no example will."
            )
        },
    )
    fim_spm_rate: Optional[float] = field(
        default=0.5,
        metadata={
            "help": (
                "Within the examples undergoing FIM transformation, this rate determines the probability "
                "of applying the Sentence Permutation Mode (SPM). "
                "Default is 0.5. A rate of 1.0 means all FIM transformations will use SPM, "
                "while a rate of 0.0 means none will."
            )
        },
    )
    truncate_or_pad: Optional[bool] = field(
        default=True,
        metadata={
            "help": (
                "Indicates whether the transformed example should be truncated or padded to maintain "
                "the same length as the original example. "
                "Default is True. If False, the function will not truncate or pad the examples."
            )
        },
    )
    fim_prefix_token: Optional[str] = field(
        default="<fim_prefix>",
        metadata={"help": ("Fill-in-Middle Prefix token. Defaults to '<fim_prefix>'.")},
    )
    fim_middle_token: Optional[str] = field(
        default="<fim_middle>",
        metadata={"help": ("Fill-in-Middle Middle token. Defaults to '<fim_middle>'.")},
    )
    fim_suffix_token: Optional[str] = field(
        default="<fim_suffix>",
        metadata={"help": ("Fill-in-Middle Suffix token. Defaults to '<fim_suffix>'.")},
    )
    pad_token: Optional[str] = field(
        default="<fim_pad>",
        metadata={
            "help": (
                "Fill-in-Middle Pad token. Used only when 'truncate_or_pad' is set to True. Defaults to '<fim_pad>'."
            )
        },
    )
    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"
        },
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    keep_linebreaks: bool = field(
        default=True, metadata={"help": "Whether to keep line breaks when using TXT files or not."}
    )

    def __post_init__(self):
        if self.streaming:
            require_version("datasets>=2.0.0", "The streaming feature requires `datasets>=2.0.0`")

        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."


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()

    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
    # information sent is the one passed as arguments along with your Python/PyTorch versions.
    send_example_telemetry("run_fim", model_args, data_args)

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )

    if training_args.should_log:
        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
        transformers.utils.logging.set_verbosity_info()

    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    # Log on each process the small summary:
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
        + f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    # Detecting last checkpoint.
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )

    # Set seed before initializing model.
    set_seed(training_args.seed)

    # Set a numpy random state for FIM transformations
    np_rng = np.random.RandomState(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 guarantee 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.
        raw_datasets = load_dataset(
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            token=model_args.token,
            streaming=data_args.streaming,
            trust_remote_code=model_args.trust_remote_code,
        )
        if "validation" not in raw_datasets.keys():
            raw_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,
                token=model_args.token,
                streaming=data_args.streaming,
                trust_remote_code=model_args.trust_remote_code,
            )
            raw_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,
                token=model_args.token,
                streaming=data_args.streaming,
                trust_remote_code=model_args.trust_remote_code,
            )
    else:
        data_files = {}
        dataset_args = {}
        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 data_args.train_file is not None
            else data_args.validation_file.split(".")[-1]
        )
        if extension == "txt":
            extension = "text"
            dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
        raw_datasets = load_dataset(
            extension,
            data_files=data_files,
            cache_dir=model_args.cache_dir,
            token=model_args.token,
            **dataset_args,
        )
        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
                token=model_args.token,
                **dataset_args,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
                token=model_args.token,
                **dataset_args,
            )

    # 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
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    config_kwargs = {
        "cache_dir": model_args.cache_dir,
        "revision": model_args.model_revision,
        "token": model_args.token,
        "trust_remote_code": model_args.trust_remote_code,
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")
        if model_args.config_overrides is not None:
            logger.info(f"Overriding config: {model_args.config_overrides}")
            config.update_from_string(model_args.config_overrides)
            logger.info(f"New config: {config}")

    tokenizer_kwargs = {
        "cache_dir": model_args.cache_dir,
        "use_fast": model_args.use_fast_tokenizer,
        "revision": model_args.model_revision,
        "token": model_args.token,
        "trust_remote_code": model_args.trust_remote_code,
    }
    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
    elif model_args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
    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.model_name_or_path:
        torch_dtype = (
            model_args.torch_dtype
            if model_args.torch_dtype in ["auto", None]
            else getattr(torch, model_args.torch_dtype)
        )
        model = AutoModelForCausalLM.from_pretrained(
            model_args.model_name_or_path,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            config=config,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            token=model_args.token,
            trust_remote_code=model_args.trust_remote_code,
            torch_dtype=torch_dtype,
            low_cpu_mem_usage=model_args.low_cpu_mem_usage,
            attn_implementation=model_args.attn_implementation,
        )

    else:
        model = AutoModelForCausalLM.from_config(
            config,
            trust_remote_code=model_args.trust_remote_code,
            attn_implementation=model_args.attn_implementation,
        )
        n_params = sum({p.data_ptr(): p.numel() for p in model.parameters()}.values())
        logger.info(f"Training new model from scratch - Total size={n_params / 2**20:.2f}M params")

    # Add the new FIM tokens to the tokenizer and resize model's vocab embeddings
    special_tokens = [data_args.fim_prefix_token, data_args.fim_middle_token, data_args.fim_suffix_token]
    if data_args.truncate_or_pad:
        special_tokens.append(data_args.pad_token)

    # Get the factor by which the embedding layer should be padded based on the device
    pad_factor = 1
    if torch.cuda.is_availble():
        pad_factor = 8

    elif is_torch_xla_available(check_is_tpu=True):
        pad_factor = 128

    # Add the new tokens to the tokenizer
    tokenizer.add_tokens(special_tokens)
    original_embeddings = model.get_input_embeddings()

    if is_deepspeed_zero3_enabled():
        import deepspeed

        with deepspeed.zero.GatheredParameters(original_embeddings.weight, modifier_rank=0):
            # Get the pre-expansion embeddings of the model and resize the embedding layer
            model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=pad_factor)
            embeddings = model.get_input_embeddings()

            # Sample the embeddings for the new tokens from a multivariate normal distribution
            # We do this so that the new embeddings are close to the original embeddings and not necessarily zero
            # More on this: https://nlp.stanford.edu/~johnhew/vocab-expansion.html
            mean = original_embeddings.mean(dim=0)
            n = original_embeddings.size()[0]
            sigma = ((original_embeddings - mean).T @ (original_embeddings - mean)) / n
            dist = torch.distributions.multivariate_normal.MultivariateNormal(
                mean,
                covariance_matrix=1e-5 * sigma,
            )
            new_token_embeddings = torch.stack(
                tuple(dist.sample() for _ in range(len(special_tokens))),
                dim=0,
            )
    else:
        original_embeddings = model.get_input_embeddings()
        # Get the pre-expansion embeddings of the model and resize the embedding layer
        model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=pad_factor)
        embeddings = model.get_input_embeddings()

        # Sample the embeddings for the new tokens from a multivariate normal distribution
        # We do this so that the new embeddings are close to the original embeddings and not necessarily zero
        # More on this: https://nlp.stanford.edu/~johnhew/vocab-expansion.html
        mean = original_embeddings.mean(dim=0)
        n = original_embeddings.size()[0]
        sigma = ((original_embeddings - mean).T @ (original_embeddings - mean)) / n
        dist = torch.distributions.multivariate_normal.MultivariateNormal(
            mean,
            covariance_matrix=1e-5 * sigma,
        )
        new_token_embeddings = torch.stack(
            tuple(dist.sample() for _ in range(len(special_tokens))),
            dim=0,
        )

    if is_deepspeed_zero3_enabled():
        import deepspeed

        with deepspeed.zero.GatheredParameters(embeddings.weight, modifier_rank=0):
            # Set the new tokens' embeddings to the newly sampled embeddings
            embeddings.weight.data[-len(special_tokens) :] = new_token_embeddings
    else:
        # Set the new tokens' embeddings to the newly sampled embeddings
        embeddings.weight.data[-len(special_tokens) :] = new_token_embeddings

    # Update the model's embeddings with the new embeddings
    model.set_input_embeddings(embeddings)

    logger.info("Added special tokens to the tokenizer and resized model's embedding layer")

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

    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")

    def tokenize_function(examples):
        with CaptureLogger(tok_logger) as cl:
            output = tokenizer(examples[text_column_name])
        # clm-fim input could be much much longer than block_size
        if "Token indices sequence length is longer than the" in cl.out:
            tok_logger.warning(
                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits"
                " before being passed to the model."
            )
        return output

    with training_args.main_process_first(desc="dataset map tokenization"):
        if not data_args.streaming:
            tokenized_datasets = raw_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,
                desc="Running tokenizer on dataset",
            )
        else:
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                remove_columns=column_names,
            )

    if data_args.block_size is None:
        block_size = tokenizer.model_max_length
        if block_size > config.max_position_embeddings:
            logger.warning(
                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
                f"Using block_size={min(1024, config.max_position_embeddings)} instead. You can change that default value by passing --block_size xxx."
            )
            block_size = min(1024, config.max_position_embeddings)
    else:
        if data_args.block_size > tokenizer.model_max_length:
            logger.warning(
                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model "
                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
            )
        block_size = min(data_args.block_size, tokenizer.model_max_length)

    # Data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
    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, and if the total_length < block_size  we exclude this batch and return an empty dict.
        # We could add padding if the model supported it instead of this drop, you can customize this part to your needs.
        total_length = (total_length // block_size) * block_size
        # Split by chunks of max_len.
        result = {
            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
            for k, t in concatenated_examples.items()
        }
        result["labels"] = result["input_ids"].copy()
        return result

    # Get the FIM-specific token ids
    prefix_tok_id = tokenizer.convert_tokens_to_ids(data_args.fim_prefix_token)
    middle_tok_id = tokenizer.convert_tokens_to_ids(data_args.fim_middle_token)
    suffix_tok_id = tokenizer.convert_tokens_to_ids(data_args.fim_suffix_token)
    pad_tok_id = None

    # If truncate_or_pad is on, also get pad token id
    if data_args.truncate_or_pad:
        pad_tok_id = tokenizer.convert_tokens_to_ids(data_args.pad_token)

    # The two functions below perform the FIM transformation on the data (either PSM or SPM or PSM+SPM)
    # Don't call fim_transform directly in .map()
    # Adapted from https://github.com/loubnabnl/santacoder-finetuning/blob/main/fim.py#L22C13-L83
    def fim_transform(example):
        """
        This function performs FIM transformation on a single example (list of tokens)
        """
        if np_rng.binomial(1, data_args.fim_rate):
            boundaries = sorted(np_rng.randint(low=0, high=len(example) + 1, size=2))

            prefix = example[: boundaries[0]]
            middle = example[boundaries[0] : boundaries[1]]
            suffix = example[boundaries[1] :]

            if data_args.truncate_or_pad:
                total_length = len(prefix) + len(middle) + len(suffix) + 3
                diff = total_length - len(example)
                if diff > 0:
                    suffix = suffix[: max(0, len(suffix) - diff)]
                elif diff < 0:
                    suffix.extend([pad_tok_id] * (-diff))

            if np_rng.binomial(1, data_args.fim_spm_rate):
                # Apply Suffix-Prefix-Middle (SPM) transformation
                transformed_example = [prefix_tok_id, suffix_tok_id] + suffix + [middle_tok_id] + prefix + middle
            else:
                # Apply Prefix-Suffix-Middle (PSM) transformation
                transformed_example = [prefix_tok_id] + prefix + [suffix_tok_id] + suffix + [middle_tok_id] + middle
        else:
            transformed_example = example

        return transformed_example

    # Below function is the one you are supposed to call in the .map() function
    def apply_fim(examples):
        """
        Apply FIM transformation to a batch of examples
        """
        fim_transform_ids = [fim_transform(ids) for ids in examples["input_ids"]]
        examples["input_ids"] = fim_transform_ids
        examples["labels"] = fim_transform_ids
        # If your application requires custom attention mask, please adjust this function's below line.
        # Since FIM transformation increases the number of tokens in input_ids and labels
        # but leaves the number of tokens unchanged in attention_masks which would cause problems
        examples["attention_mask"] = [[1] * len(mask) for mask in examples["input_ids"]]
        return examples

    # 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/process#map

    # FIM transformations are only supposed to be applied before group_texts processing otherwise some sentences will
    # have 3-4 more tokens than others due to probabilistic addition of FIM-specific tokens which will raise errors
    with training_args.main_process_first(desc="processing texts together"):
        if not data_args.streaming:
            fim_datasets = tokenized_datasets.map(
                apply_fim,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Performing FIM transformation",
            )
            lm_datasets = fim_datasets.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc=f"Grouping texts in chunks of {block_size}",
            )
        else:
            fim_datasets = tokenized_datasets.map(
                apply_fim,
                batched=True,
            )
            lm_datasets = fim_datasets.map(
                group_texts,
                batched=True,
            )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = lm_datasets["train"]
        if data_args.max_train_samples is not None:
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = lm_datasets["validation"]
        if data_args.max_eval_samples is not None:
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset = eval_dataset.select(range(max_eval_samples))

        def preprocess_logits_for_metrics(logits, labels):
            if isinstance(logits, tuple):
                # Depending on the model and config, logits may contain extra tensors,
                # like past_key_values, but logits always come first
                logits = logits[0]
            return logits.argmax(dim=-1)

        metric = evaluate.load("accuracy")

        def compute_metrics(eval_preds):
            preds, labels = eval_preds
            # preds have the same shape as the labels, after the argmax(-1) has been calculated
            # by preprocess_logits_for_metrics but we need to shift the labels
            labels = labels[:, 1:].reshape(-1)
            preds = preds[:, :-1].reshape(-1)
            return metric.compute(predictions=preds, references=labels)

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset if training_args.do_train else None,
        eval_dataset=eval_dataset if training_args.do_eval else None,
        processing_class=tokenizer,
        # Data collator will default to DataCollatorWithPadding, so we change it.
        data_collator=default_data_collator,
        compute_metrics=compute_metrics
        if training_args.do_eval and not is_torch_xla_available(check_is_tpu=True)
        else None,
        preprocess_logits_for_metrics=(
            preprocess_logits_for_metrics
            if training_args.do_eval and not is_torch_xla_available(check_is_tpu=True)
            else None
        ),
    )

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        trainer.save_model()  # Saves the tokenizer too for easy upload

        metrics = train_result.metrics

        max_train_samples = (
            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))

        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        metrics = trainer.evaluate()

        max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
        try:
            perplexity = math.exp(metrics["eval_loss"])
        except OverflowError:
            perplexity = float("inf")
        metrics["perplexity"] = perplexity

        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "text-generation"}
    if data_args.dataset_name is not None:
        kwargs["dataset_tags"] = data_args.dataset_name
        if data_args.dataset_config_name is not None:
            kwargs["dataset_args"] = data_args.dataset_config_name
            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
        else:
            kwargs["dataset"] = data_args.dataset_name

    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)


def _mp_fn(index):
    # For xla_spawn (TPUs)
    main()


if __name__ == "__main__":
    main()