train.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 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 (GPT, GPT-2, CTRL, ...)
on a text file or a dataset without using HuggingFace Trainer.

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 can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
import sys
import time
sys.path.insert(1, 'src')

_TRAIN_START_TIME = time.time()

import argparse
import json
import logging
import math
import os
import random
from itertools import chain
from pathlib import Path

import datasets
import torch
from accelerate import Accelerator, DistributedType
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from datasets import load_dataset
from huggingface_hub import Repository, create_repo
from torch.utils.data import DataLoader
from tqdm.auto import tqdm

import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_MAPPING,
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    SchedulerType,
    default_data_collator,
    get_scheduler,
)
from utils.cache_utils import DynamicCache
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version

from utils.global_vars import get_args, set_args
from peft import LoraConfig, get_peft_model

from llama import LlamaConfig, LlamaForCausalLM

# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
# check_min_version("4.36.0")

logger = get_logger(__name__)

require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")

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

import datetime
# torch.distributed.init_process_group(backend="nccl", timeout=datetime.timedelta(seconds=108000))


def parse_args():
    parser = argparse.ArgumentParser(description="Finetune a transformers model on a causal language modeling task")

    parser.add_argument('--method', choices=['hh', 'conv'])
    parser.add_argument('--mem_size', type=int, default=128)
    parser.add_argument('--local_len', type=int, default=0)
    parser.add_argument('--hh_keep_rate', type=float, default=0.5)
    parser.add_argument('--n_convlayer', type=int, default=1)
    parser.add_argument('--kernel_size', type=int, default=21)
    parser.add_argument('--hidden_act', default='relu')
    parser.add_argument('--expand', type=int, default=1)

    parser.add_argument('--eval_interval', type=int, default=20)
    parser.add_argument('--eval_iter', type=int)
    parser.add_argument('--stream_tokenizer', action='store_true')
    parser.add_argument('--normalizer_init', type=float, default=0.5)
    parser.add_argument('--rope_change', action='store_true')

    parser.add_argument('--lora_finetuning', action='store_true')
    parser.add_argument('--memory_lr_scale', type=float, default=1)
    parser.add_argument('--norm_lr_scale', type=float, default=1)
    parser.add_argument('--lora_lr_scale', type=float, default=1)

    parser.add_argument('--exit-after-steps', type=int)
    parser.add_argument('--auto_resume', action='store_true')
    parser.add_argument('--checkpoint_duration_in_mins', type=int, default=215)

    parser.add_argument(
        "--dataset_name",
        type=str,
        default=None,
        help="The name of the dataset to use (via the datasets library).",
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The configuration name of the dataset to use (via the datasets library).",
    )
    parser.add_argument(
        "--train_file", type=str, default=None, help="A csv, txt or a json file containing the training data."
    )
    parser.add_argument(
        "--validation_file", type=str, default=None, help="A csv, txt or a json file containing the validation data."
    )
    parser.add_argument(
        "--validation_split_percentage",
        default=5,
        help="The percentage of the train set used as validation set in case there's no validation split",
    )
    parser.add_argument(
        "--model_name_or_path",
        type=str,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
        required=False,
    )
    parser.add_argument(
        "--config_name",
        type=str,
        default=None,
        help="Pretrained config name or path if not the same as model_name",
    )
    parser.add_argument(
        "--tokenizer_name",
        type=str,
        default=None,
        help="Pretrained tokenizer name or path if not the same as model_name",
    )
    parser.add_argument(
        "--use_slow_tokenizer",
        action="store_true",
        help="If passed, will use a slow tokenizer (not backed by the 🤗 Tokenizers library).",
    )
    parser.add_argument(
        "--per_device_train_batch_size",
        type=int,
        default=1,
        help="Batch size (per device) for the training dataloader.",
    )
    parser.add_argument(
        "--per_device_eval_batch_size",
        type=int,
        default=1,
        help="Batch size (per device) for the evaluation dataloader.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=5e-5,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument("--weight_decay", type=float, default=0.0, help="Weight decay to use.")
    parser.add_argument("--num_train_epochs", type=int, default=3, help="Total number of training epochs to perform.")
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--lr_scheduler_type",
        type=SchedulerType,
        default="linear",
        help="The scheduler type to use.",
        choices=["linear", "cosine", "cosine_with_restarts", "polynomial", "constant", "constant_with_warmup"],
    )
    parser.add_argument(
        "--num_warmup_steps", type=int, default=0, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument("--output_dir", type=str, default=None, help="Where to store the final model.")
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--model_type",
        type=str,
        default=None,
        help="Model type to use if training from scratch.",
        choices=MODEL_TYPES,
    )
    parser.add_argument(
        '--clean_period', type=int
    )
    parser.add_argument(
        "--block_size",
        type=int,
        default=None,
        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)."
        ),
    )
    parser.add_argument(
        "--preprocessing_num_workers",
        type=int,
        default=None,
        help="The number of processes to use for the preprocessing.",
    )
    parser.add_argument(
        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
    )
    parser.add_argument(
        "--no_keep_linebreaks", action="store_true", help="Do not keep line breaks when using TXT files."
    )
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument(
        "--hub_model_id", type=str, help="The name of the repository to keep in sync with the local `output_dir`."
    )
    parser.add_argument("--hub_token", type=str, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--trust_remote_code",
        type=bool,
        default=False,
        help=(
            "Whether or not to allow for custom models defined on the Hub in their own modeling files. 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."
        ),
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=str,
        default=None,
        help="Whether the various states should be saved at the end of every n steps, or 'epoch' for each epoch.",
    )
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help="If the training should continue from a checkpoint folder.",
    )
    parser.add_argument(
        "--with_tracking",
        action="store_true",
        help="Whether to enable experiment trackers for logging.",
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="all",
        help=(
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`,'
            ' `"wandb"`, `"comet_ml"` and `"clearml"`. Use `"all"` (default) to report to all integrations. '
            "Only applicable when `--with_tracking` is passed."
        ),
    )
    parser.add_argument(
        "--low_cpu_mem_usage",
        action="store_true",
        help=(
            "It is an option to create the model as an empty shell, then only materialize its parameters when the pretrained weights are loaded. "
            "If passed, LLM loading time and RAM consumption will be benefited."
        ),
    )
    args = parser.parse_args()

    # Sanity checks
    if args.dataset_name is None and args.train_file is None and args.validation_file is None:
        raise ValueError("Need either a dataset name or a training/validation file.")
    else:
        if args.train_file is not None:
            extension = args.train_file.split(".")[-1]
            if extension not in ["csv", "json", "txt"]:
                raise ValueError("`train_file` should be a csv, json or txt file.")
        if args.validation_file is not None:
            extension = args.validation_file.split(".")[-1]
            if extension not in ["csv", "json", "txt"]:
                raise ValueError("`validation_file` should be a csv, json or txt file.")

    if args.push_to_hub:
        if args.output_dir is None:
            raise ValueError("Need an `output_dir` to create a repo when `--push_to_hub` is passed.")
    
    set_args(args)

    return args


def main():
    args = parse_args()
    
    print(args)

    # 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_clm_no_trainer", args)

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    # If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
    # in the environment
    accelerator_log_kwargs = {}

    if args.with_tracking:
        accelerator_log_kwargs["log_with"] = args.report_to
        accelerator_log_kwargs["project_dir"] = args.output_dir

    accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps, **accelerator_log_kwargs)

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.push_to_hub:
            # Retrieve of infer repo_name
            repo_name = args.hub_model_id
            if repo_name is None:
                repo_name = Path(args.output_dir).absolute().name
            # Create repo and retrieve repo_id
            repo_id = create_repo(repo_name, exist_ok=True, token=args.hub_token).repo_id
            # Clone repo locally
            repo = Repository(args.output_dir, clone_from=repo_id, token=args.hub_token)

            with open(os.path.join(args.output_dir, ".gitignore"), "w+") as gitignore:
                if "step_*" not in gitignore:
                    gitignore.write("step_*\n")
                if "epoch_*" not in gitignore:
                    gitignore.write("epoch_*\n")
        elif args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)
    accelerator.wait_for_everyone()

    # 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 args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
            )
    else:
        data_files = {}
        dataset_args = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
        extension = args.train_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
            dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
        raw_datasets = load_dataset(extension, data_files=data_files, **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[:{args.validation_split_percentage}%]",
                **dataset_args,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{args.validation_split_percentage}%:]",
                **dataset_args,
            )
    
    if args.stream_tokenizer:
        # data streaming
        args.dataset_num_shards = 1024
        raw_datasets = datasets.IterableDatasetDict({
            "train": raw_datasets["train"].to_iterable_dataset(num_shards=args.dataset_num_shards),
            "validation": raw_datasets["validation"].to_iterable_dataset(num_shards=args.dataset_num_shards),
        })

    # 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.

    # Load pretrained model and tokenizer
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    if args.config_name:
        config = AutoConfig.from_pretrained(
            args.config_name,
            trust_remote_code=args.trust_remote_code,
        )
    elif args.model_name_or_path:
        config = LlamaConfig.from_pretrained(
            args.model_name_or_path,
            trust_remote_code=args.trust_remote_code,
            attn_implementation="eager",
        )
    else:
        config = CONFIG_MAPPING[args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")

    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            args.tokenizer_name, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
        )
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
        )
    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 args.model_name_or_path:

        if args.rope_change:
            if accelerator.is_main_process:
                print('Change Rope Configuration: type: Linear, factor: %f'%(args.block_size * args.clean_period/config.max_position_embeddings))
            config.rope_scaling = {'type': 'linear', 
                'factor': (args.block_size * args.clean_period/config.max_position_embeddings)}

        if args.lora_finetuning:
            base_model = LlamaForCausalLM.from_pretrained(
                args.model_name_or_path,
                from_tf=bool(".ckpt" in args.model_name_or_path),
                config=config,
                low_cpu_mem_usage=args.low_cpu_mem_usage,
                trust_remote_code=args.trust_remote_code,
                attn_implementation="eager"
            )
            config = LoraConfig(
                r=8,
                lora_alpha=16,
                target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
                lora_dropout=0,
                bias="none",
                task_type="CAUSAL_LM",
            )
            model = get_peft_model(base_model, config)

        else:
            base_model = None
            model = LlamaForCausalLM.from_pretrained(
                args.model_name_or_path,
                from_tf=bool(".ckpt" in args.model_name_or_path),
                config=config,
                low_cpu_mem_usage=args.low_cpu_mem_usage,
                trust_remote_code=args.trust_remote_code,
                attn_implementation="eager"
            )

    else:
        raise NotImplementedError
        logger.info("Training new model from scratch")
        model = AutoModelForCausalLM.from_config(config, trust_remote_code=args.trust_remote_code)

    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
    # on a small vocab and want a smaller embedding size, remove this test.
    embedding_size = model.get_input_embeddings().weight.shape[0]
    if len(tokenizer) > embedding_size:
        model.resize_token_embeddings(len(tokenizer))

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

    def tokenize_function(examples):
        return tokenizer(examples[text_column_name])

    if args.stream_tokenizer:
        with accelerator.main_process_first():
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                remove_columns=column_names,
            )
    else:
        with accelerator.main_process_first():
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=True,
                num_proc=args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not args.overwrite_cache,
                desc="Running tokenizer on dataset",
            )
    
    block_size = args.block_size * args.clean_period

    if args.clean_period != 1:
        print(f'Per sample length {block_size}, can be truncated into {args.clean_period} parts.')
    assert block_size < tokenizer.model_max_length

    # if 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 args.block_size > tokenizer.model_max_length:
    #         logger.warning(
    #             f"The block_size passed ({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(args.block_size, tokenizer.model_max_length)

    # Main 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

    # 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

    if args.stream_tokenizer:
        with accelerator.main_process_first():
            lm_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
            )
    else:
        with accelerator.main_process_first():
            lm_datasets = tokenized_datasets.map(
                group_texts,
                batched=True,
                num_proc=args.preprocessing_num_workers,
                load_from_cache_file=not args.overwrite_cache,
                desc=f"Grouping texts in chunks of {block_size}",
            )

    train_dataset = lm_datasets["train"]
    eval_dataset = lm_datasets["validation"]

    # # Log a few random samples from the training set:
    # for index in random.sample(range(len(train_dataset)), 3):
    #     logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset.shuffle(), collate_fn=default_data_collator, batch_size=args.per_device_train_batch_size
    )
    eval_dataloader = DataLoader(
        eval_dataset, collate_fn=default_data_collator, batch_size=args.per_device_eval_batch_size
    )

    # customize optimzer
    no_decay = ["bias", "batch_norm.weight"]
    memory_params_nd, memory_params, norm_params, lora_params = [], [], [], []
    for n, p in model.named_parameters():
        if 'memory_saver' in n and not any(nd in n for nd in no_decay):
            p.requires_grad_(True)
            memory_params.append(p)
            logger.info('Trainable Parameter: %s, lr scale %f, weight_decay'%(n, args.memory_lr_scale))
        elif 'memory_saver' in n and any(nd in n for nd in no_decay):
            p.requires_grad_(True)
            memory_params_nd.append(p)
            logger.info('Trainable Parameter: %s, lr scale %f, no weight_decay'%(n, args.memory_lr_scale))
        elif 'embed' in n or 'norm' in n:
            p.requires_grad_(True)
            norm_params.append(p)
            logger.info('Trainable Parameter: %s, lr scale %f, weight_decay'%(n, args.norm_lr_scale))
        elif 'lora' in n:
            p.requires_grad_(True)
            lora_params.append(p)
            logger.info('Trainable Parameter: %s, lr scale %f, weight_decay'%(n, args.lora_lr_scale))
        else:
            p.requires_grad_(False)
    
    if args.lora_finetuning:
        model.print_trainable_parameters()

    optimizer_grouped_parameters = [
        {"params": memory_params_nd, "weight_decay": 0.0, "lr": args.learning_rate * args.memory_lr_scale},
        {"params": memory_params, "weight_decay": args.weight_decay, "lr": args.learning_rate * args.memory_lr_scale},
        {"params": norm_params, "weight_decay": args.weight_decay, "lr": args.learning_rate * args.norm_lr_scale},
        {"params": lora_params, "weight_decay": args.weight_decay, "lr": args.learning_rate * args.lora_lr_scale},
    ]

    # Optimizer
    # Split weights in two groups, one with weight decay and the other not.
    optimizer = torch.optim.AdamW(optimizer_grouped_parameters)

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    # num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    # if args.max_train_steps is None:
    #     args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    #     overrode_max_train_steps = True

    if not args.lora_finetuning and args.method=='hh':
        model, optimizer, train_dataloader, eval_dataloader = accelerator.prepare(
            model, optimizer, train_dataloader, eval_dataloader
        )
        evaluate(model, eval_dataloader, accelerator, logger, 0, verbose=True)
        sys.exit()

    assert args.max_train_steps is not None

    lr_scheduler = get_scheduler(
        name=args.lr_scheduler_type,
        optimizer=optimizer,
        num_warmup_steps=args.num_warmup_steps * accelerator.num_processes,
        num_training_steps=args.max_train_steps * accelerator.num_processes,
    ) 
    # NOTE added by Ruisi: please check `accelerator.num_processes`. might needs to modify to args.gradient_accumulation_steps. 
    # (depend on whether using accelerate package)

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
    )

    # if base_model is not None:
    #     accelerator.register_for_checkpointing(base_model)

    # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
    if accelerator.distributed_type == DistributedType.TPU:
        model.tie_weights()

    # # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    # num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    # if overrode_max_train_steps:
    #     args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # # Afterwards we recalculate our number of training epochs
    # args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # Figure out how many steps we should save the Accelerator states
    checkpointing_steps = args.checkpointing_steps
    if checkpointing_steps is not None and checkpointing_steps.isdigit():
        checkpointing_steps = int(checkpointing_steps)
    checkpointing_steps = None

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if args.with_tracking:
        experiment_config = vars(args)
        # TensorBoard cannot log Enums, need the raw value
        experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
        accelerator.init_trackers("clm_no_trainer", experiment_config)

    # Train!
    total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    # logger.info(f"  Num examples = {len(train_dataset)}")
    # logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {int(total_batch_size/args.clean_period)}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    # Only show the progress bar once on each machine.
    # progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
    completed_steps = 0
    starting_epoch = 0
    exit_checkpointing_flag = False

    # Potentially load in the weights and states from a previous save
    if args.auto_resume:
        ckpt_dir = os.path.join(args.output_dir)
        checkpoint_names = []
        if os.path.isdir(ckpt_dir):
            is_ckpt_folder = lambda fn: fn.startswith('step_')
            checkpoint_names = [fn for fn in os.listdir(ckpt_dir) if is_ckpt_folder(fn)]
        if len(checkpoint_names) > 0:
            checkpoint_names = sorted(checkpoint_names, key=lambda p: int(p.split('_')[1]))
            args.resume_from_checkpoint = os.path.join(ckpt_dir, checkpoint_names[-1])
            logger.info('Detect checkpoint from: %s'%args.resume_from_checkpoint, main_process_only=True)
            completed_steps = int(checkpoint_names[-1].split("_")[1])
        else:
            logger.info('No checkpoint detected: %s'%ckpt_dir, main_process_only=True)
            args.resume_from_checkpoint = None


    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
            checkpoint_path = args.resume_from_checkpoint
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
            dirs.sort(key=os.path.getctime)
            path = dirs[-1]  # Sorts folders by date modified, most recent checkpoint is the last
            checkpoint_path = path
            path = os.path.basename(checkpoint_path)

        accelerator.print(f"Resumed from checkpoint: {checkpoint_path}")
        accelerator.load_state(checkpoint_path)
        # Extract `epoch_{i}` or `step_{i}`
        training_difference = os.path.splitext(path)[0]


    # update the progress_bar if load from checkpoint
    # progress_bar.update(completed_steps)

    while completed_steps < args.max_train_steps:
        model.train()
        if args.with_tracking:
            total_loss = 0
        if args.resume_from_checkpoint:
            # We skip the first `n` batches in the dataloader when resuming from a checkpoint
            active_dataloader = accelerator.skip_first_batches(train_dataloader, completed_steps)
        else:
            active_dataloader = train_dataloader
            
        losses = []
        for step, batch in enumerate(active_dataloader):

            past_key_values = DynamicCache() 
            batch_loss = {'lm': []}
            for seq_idx in range(args.clean_period):
                
                with accelerator.accumulate(model):

                    batch_item = {}
                    batch_item['use_cache'] = True
                    start, end = seq_idx * args.block_size, (seq_idx+1) * args.block_size
                    
                    for key in batch:
                        key_start = 0 if key=='attention_mask' and batch_item['use_cache'] else start
                        batch_item[key] = batch[key][:, key_start: end]

                    batch_item['past_key_values'] = past_key_values
                    
                    outputs = model(**batch_item)
                    loss = outputs.loss
                    past_key_values = outputs.past_key_values
                    # We keep track of the loss at each epoch
                    if args.with_tracking:
                        total_loss += loss.detach().float()
                    batch_loss['lm'].append(loss.item())
                    accelerator.backward(loss)
                    optimizer.step()
                    lr_scheduler.step()
                    optimizer.zero_grad()
                loss = torch.tensor(batch_loss['lm'], device=loss.device).mean()
                losses.append(accelerator.gather(loss.repeat(args.per_device_train_batch_size)))

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                # progress_bar.update(1)
                completed_steps += 1
                if completed_steps % args.eval_interval == 0:
                    evaluate(model, eval_dataloader, accelerator, logger, completed_steps)
                
                losses = torch.cat(losses)
                try:
                    train_loss = torch.mean(losses)
                    train_perplexity = math.exp(train_loss)
                except OverflowError:
                    train_perplexity = float("inf")

                logger.info(f"step {completed_steps}: perplexity: {train_perplexity} eval_loss: {train_loss}")
                losses = []

                if isinstance(checkpointing_steps, int):
                    if completed_steps % checkpointing_steps == 0:
                        # accelerator.wait_for_everyone()
                        # unwrapped_model = accelerator.unwrap_model(model)
                        # unwrapped_model.save_pretrained(
                        #     args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
                        # )

                        output_dir = f"step_{completed_steps}"
                        if args.output_dir is not None:
                            output_dir = os.path.join(args.output_dir, output_dir)
                        accelerator.save_state(output_dir)
                
                if not exit_checkpointing_flag and args.checkpoint_duration_in_mins:
                    train_time = (time.time() - _TRAIN_START_TIME) / 60.0
                    done_cuda = torch.cuda.IntTensor(
                        [train_time > args.checkpoint_duration_in_mins])
                    torch.distributed.all_reduce(
                        done_cuda, op=torch.distributed.ReduceOp.MAX)
                    done = done_cuda.item()
                    if done:
                        output_dir = f"step_{completed_steps}"
                        if args.output_dir is not None:
                            output_dir = os.path.join(args.output_dir, output_dir)
                        accelerator.save_state(output_dir)
                        logger.info('checkpointing program after {} minutes'.format(train_time))
                        exit_checkpointing_flag = True
                        sys.exit()
            
                if args.exit_after_steps is not None and completed_steps % args.exit_after_steps == 0:
                    sys.exit()

                if completed_steps >= args.max_train_steps:
                    break

        if args.with_tracking:
            accelerator.log(
                {
                    "perplexity": perplexity,
                    "eval_loss": eval_loss,
                    "train_loss": total_loss.item() / len(train_dataloader),
                    "epoch": epoch,
                    "step": completed_steps,
                },
                step=completed_steps,
            )

        if args.push_to_hub and epoch < args.num_train_epochs - 1:
            accelerator.wait_for_everyone()
            unwrapped_model = accelerator.unwrap_model(model)
            unwrapped_model.save_pretrained(
                args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
            )
            if accelerator.is_main_process:
                tokenizer.save_pretrained(args.output_dir)
                repo.push_to_hub(
                    commit_message=f"Training in progress epoch {epoch}", blocking=False, auto_lfs_prune=True
                )

        if args.checkpointing_steps == "epoch":
            output_dir = f"epoch_{epoch}"
            if args.output_dir is not None:
                output_dir = os.path.join(args.output_dir, output_dir)
            accelerator.save_state(output_dir)

    if args.with_tracking:
        accelerator.end_training()

    if args.output_dir is not None:
        accelerator.wait_for_everyone()
        unwrapped_model = accelerator.unwrap_model(model)
        if args.lora_finetuning:
            unwrapped_model = unwrapped_model.merge_and_unload()
        unwrapped_model.save_pretrained(
            args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
        )
        if accelerator.is_main_process:
            tokenizer.save_pretrained(args.output_dir)
            if args.push_to_hub:
                repo.push_to_hub(commit_message="End of training", auto_lfs_prune=True)

            # with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
            #     json.dump({"perplexity": perplexity}, f)


def print_str(str, file, accelerator):
    logger.info(str)
    if accelerator.is_main_process:
        with open(file, 'a') as f:
            print(str, file=f)


def evaluate(model, eval_dataloader, accelerator, logger, completed_step, verbose=False):

    args = get_args()

    model.eval()
    losses = []
    for step, batch in enumerate(eval_dataloader):
        past_key_values = DynamicCache() 
        batch_loss = {'lm': []}
        if verbose:
            if accelerator.is_main_process:
                print(f'Evaluating Step {step} / {args.eval_iter if args.eval_iter is not None else 0}')

        with torch.no_grad():
            for seq_idx in range(args.clean_period):

                batch_item = {}
                batch_item['use_cache'] = True
                start, end = seq_idx * args.block_size, (seq_idx+1) * args.block_size
                
                for key in batch:
                    key_start = 0 if key=='attention_mask' and batch_item['use_cache'] else start
                    batch_item[key] = batch[key][:, key_start: end]

                batch_item['past_key_values'] = past_key_values 
                outputs = model(**batch_item)

                loss = outputs.loss
                past_key_values = outputs.past_key_values
                batch_loss['lm'].append(loss.item())

        loss = torch.tensor(batch_loss['lm'], device=loss.device).mean()
        # losses.append(accelerator.gather_for_metrics(loss.repeat(args.per_device_eval_batch_size)))
        losses.append(accelerator.gather(loss.repeat(args.per_device_eval_batch_size)))

        if args.eval_iter is not None and step==args.eval_iter:
            break

        if verbose and step % 200 ==0 :
            losses_temp = torch.cat(losses)
            try:
                eval_loss = torch.mean(losses_temp)
                perplexity = math.exp(eval_loss)
            except OverflowError:
                perplexity = float("inf")

            logger.info(f"Evaluation: step {completed_step}: perplexity: {perplexity} eval_loss: {eval_loss}")

    losses = torch.cat(losses)
    try:
        eval_loss = torch.mean(losses)
        perplexity = math.exp(eval_loss)
    except OverflowError:
        perplexity = float("inf")

    logger.info(f"Evaluation: step {completed_step}: perplexity: {perplexity} eval_loss: {eval_loss}")


if __name__ == "__main__":
    main()