incr_pretrain_tinyllama.py

# Modification by Light Transport Entertainment Inc.

# Copyright 2020 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.

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 logging
import numpy as np
import math
import os
import sys
from dataclasses import dataclass, field
from itertools import chain
from typing import Optional, List, Dict, Any, Mapping
from pathlib import Path
import datasets
import torch

import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_FOR_CAUSAL_LM_MAPPING,
    AutoConfig,
    AutoModelForCausalLM,
    LlamaForCausalLM,
    LlamaTokenizer,
    AutoTokenizer,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    is_torch_tpu_available,
    set_seed,
    BitsAndBytesConfig
)
from transformers.testing_utils import CaptureLogger
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import send_example_telemetry
from transformers.utils.versions import require_version

from sklearn.metrics import accuracy_score
from peft import LoraConfig, TaskType, get_peft_model, PeftModel, get_peft_model_state_dict
from peft.tuners.lora import LoraLayer
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
from datasets import load_dataset, concatenate_datasets
from tokenizers import Tokenizer

jp_vocab_size = 51470

#def get_embed_size(nvocab, m = 64):
#    return m * (round(nvocab / m) + 1)

class SavePeftModelCallback(transformers.TrainerCallback):
    def save_model(self, args, state, kwargs):
        if state.best_model_checkpoint is not None:
            checkpoint_folder = os.path.join(state.best_model_checkpoint, "pt_lora_model")
        else:
            checkpoint_folder = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")

        peft_model_path = os.path.join(checkpoint_folder, "pt_lora_model")
        kwargs["model"].save_pretrained(peft_model_path)
        kwargs["tokenizer"].save_pretrained(peft_model_path)

    def on_save(self, args, state, control, **kwargs):
        self.save_model(args, state, kwargs)
        return control

    def on_train_end(self, args, state, control, **kwargs):
        peft_model_path = os.path.join(args.output_dir, "pt_lora_model")
        kwargs["model"].save_pretrained(peft_model_path)
        kwargs["tokenizer"].save_pretrained(peft_model_path)


def prepare_model_for_kbit_training(model, use_gradient_checkpointing=True):
    r"""
    This method wraps the entire protocol for preparing a model before running a training. This includes:
        1- Cast the layernorm in fp32 2- making output embedding layer require grads 3- Add the upcasting of the lm
        head to fp32

    Args:
        model, (`transformers.PreTrainedModel`):
            The loaded model from `transformers`
    """
    loaded_in_kbit = getattr(model, "is_loaded_in_8bit", False) or getattr(model, "is_loaded_in_4bit", False)

    for name, param in model.named_parameters():
        # freeze base model's layers
        param.requires_grad = False

    # cast all non INT8/INT4 parameters to fp32
    for param in model.parameters():
        if ((param.dtype == torch.float16) or (param.dtype == torch.bfloat16)) and loaded_in_kbit:
            param.data = param.data.to(torch.float32)

    for name, module in model.named_modules():
        if 'norm' in name:
            module = module.to(torch.float32)

    if loaded_in_kbit and use_gradient_checkpointing:
        # For backward compatibility
        if hasattr(model, "enable_input_require_grads"):
            model.enable_input_require_grads()
        else:
            def make_inputs_require_grad(module, _input, output):
                output.requires_grad_(True)

            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
        # enable gradient checkpointing for memory efficiency
        model.gradient_checkpointing_enable()

    return model


def accuracy(predictions, references, normalize=True, sample_weight=None):
        return {
            "accuracy": float(
                accuracy_score(references, predictions, normalize=normalize, sample_weight=sample_weight)
            )
        }


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 accuracy(predictions=preds, references=labels)


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)


def fault_tolerance_data_collator(features: List) -> Dict[str, Any]:
    if not isinstance(features[0], Mapping):
        features = [vars(f) for f in features]
    first = features[0]
    batch = {}

    # Special handling for labels.
    # Ensure that tensor is created with the correct type
    # (it should be automatically the case, but let's make sure of it.)
    if "label" in first and first["label"] is not None:
        label = first["label"].item() if isinstance(first["label"], torch.Tensor) else first["label"]
        dtype = torch.long if isinstance(label, int) else torch.float
        batch["labels"] = torch.tensor([f["label"] for f in features], dtype=dtype)
    elif "label_ids" in first and first["label_ids"] is not None:
        if isinstance(first["label_ids"], torch.Tensor):
            batch["labels"] = torch.stack([f["label_ids"] for f in features])
        else:
            dtype = torch.long if isinstance(first["label_ids"][0], int) else torch.float
            batch["labels"] = torch.tensor([f["label_ids"] for f in features], dtype=dtype)

    # Handling of all other possible keys.
    # Again, we will use the first element to figure out which key/values are not None for this model.

    try:
        for k, v in first.items():
            if k not in ("label", "label_ids") and v is not None and not isinstance(v, str):
                if isinstance(v, torch.Tensor):
                    batch[k] = torch.stack([f[k] for f in features])
                elif isinstance(v, np.ndarray):
                    batch[k] = torch.tensor(np.stack([f[k] for f in features]))
                else:
                    batch[k] = torch.tensor([f[k] for f in features])
    except ValueError: # quick fix by simply take the first example
        for k, v in first.items():
            if k not in ("label", "label_ids") and v is not None and not isinstance(v, str):
                if isinstance(v, torch.Tensor):
                    batch[k] = torch.stack([features[0][k]] * len(features))
                elif isinstance(v, np.ndarray):
                    batch[k] = torch.tensor(np.stack([features[0][k]] * len(features)))
                else:
                    batch[k] = torch.tensor([features[0][k]] * len(features))

    return batch


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."
            )
        },
    )
    tokenizer_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The tokenizer 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)."},
    )
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": (
                "Will use the token generated when running `huggingface-cli login` (necessary to use this script "
                "with private models)."
            )
        },
    )
    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"],
        },
    )

    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_dir: 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)."
            )
        },
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    validation_split_percentage: Optional[float] = field(
        default=0.05,
        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."}
    )
    data_cache_dir: Optional[str] = field(default="./", metadata={"help": "The datasets processed stored"})

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


@dataclass
class MyTrainingArguments(TrainingArguments):
    trainable : Optional[str] = field(default="q_proj,v_proj")
    lora_rank : Optional[int] = field(default=8)
    lora_dropout : Optional[float] = field(default=0.1)
    lora_alpha : Optional[float] = field(default=32.)
    modules_to_save : Optional[str] = field(default=None)
    debug_mode : Optional[bool] = field(default=False)
    peft_path : Optional[str] = field(default=None)
    flash_attn : Optional[bool] = field(default=False)
    double_quant: Optional[bool] = field(default=True)
    quant_type: Optional[str] = field(default="nf4")
    load_in_kbits: Optional[int] = field(default=16)


logger = logging.getLogger(__name__)


def main():

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, MyTrainingArguments))
    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 training_args.flash_attn:
    #    from flash_attn_patch import replace_llama_attn_with_flash_attn
    #    replace_llama_attn_with_flash_attn()

    # 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", model_args, data_args)

    # Setup logging
    logging.basicConfig(format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,  # if training_args.local_rank in [-1, 0] else logging.WARN,
        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()
    # transformers.tokenization_utils.logging.set_verbosity_warning()

    # 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: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
    )

    # 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)

    config_kwargs = {
        "cache_dir": model_args.cache_dir,
        "revision": model_args.model_revision,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    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,
        "padding": True,
        "use_auth_token": True if model_args.use_auth_token else None,
    }
    #if model_args.tokenizer_name:
    #    tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
    if model_args.tokenizer_name_or_path:
        tokenizer = LlamaTokenizer.from_pretrained(model_args.tokenizer_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."
        )
    tokenizer.add_eos_token = True

    # Preprocessing the datasets.
    # First we tokenize all the texts.
    # 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"])
        # clm 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

    if data_args.block_size is None:
        block_size = tokenizer.model_max_length
        if block_size > 1024:
            logger.warning(
                "The chosen tokenizer supports a `model_max_length` that is longer than the default `block_size` value"
                " of 1024. If you would like to use a longer `block_size` up to `tokenizer.model_max_length` you can"
                " override this default with `--block_size xxx`."
            )
            block_size = 1024
    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)

    # 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.
        #print(examples.keys())
        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 >= block_size:
            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

    with training_args.main_process_first(desc="dataset map tokenization and grouping"):

        filename="charshu"

        cache_path = os.path.join(data_args.data_cache_dir, filename+f"_{block_size}")
        os.makedirs(cache_path, exist_ok=True)

        try:
            processed_dataset = datasets.load_from_disk(cache_path, keep_in_memory=False)
            logger.info(f'training datasets-{filename} has been loaded from disk')

            print(processed_dataset)

        except Exception:

            raw_dataset = load_dataset(data_args.dataset_dir)

            # only use 'text' columns
            # remove 'lm_score' columns(float), which causes 'float object is not iterable' error in 'group_texts'
            for split_name in raw_dataset.keys():
                raw_dataset[split_name] = raw_dataset[split_name].select_columns(['text'])

            print(raw_dataset)


            cache_dir = os.path.join(data_args.data_cache_dir, filename+f"_text_{block_size}")
            os.makedirs(cache_dir, exist_ok=True)

            for split_name in raw_dataset.keys():
                split_cache_dir = os.path.join(cache_dir, split_name)
                os.makedirs(split_cache_dir, exist_ok=True)
            
            tokenized_dataset = raw_dataset.map(
                tokenize_function,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns="text",
                load_from_cache_file=True,
                keep_in_memory=False,
                cache_file_names = {split: os.path.join(os.path.join(cache_dir, split), 'tokenized.arrow') for split in raw_dataset},
                desc="Running tokenizer on dataset",
            )

            print(tokenized_dataset)
            grouped_datasets = tokenized_dataset.map(
                group_texts,
                batched=True,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=True,
                keep_in_memory=False,
                cache_file_names = {split: os.path.join(os.path.join(cache_dir, split), 'grouped.arrow') for split in tokenized_dataset},
                desc=f"Grouping texts in chunks of {block_size}",
            )
            processed_dataset = grouped_datasets
            print(processed_dataset)
            processed_dataset.save_to_disk(cache_path)

        lm_datasets = processed_dataset # splits = 'train', 'validation', 'test'

    if training_args.do_train:
        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))
        logger.info(f"Num train_samples  {len(train_dataset)}")
        logger.info("Training example:")
        logger.info(tokenizer.decode(train_dataset[0]['input_ids']))
    if training_args.do_eval:
        eval_dataset = lm_datasets["test"]
        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))
        logger.info(f"Num eval_samples  {len(eval_dataset)}")
        logger.info("Evaluation example:")
        logger.info(tokenizer.decode(eval_dataset[0]['input_ids']))
    compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
    if training_args.load_in_kbits in [4, 8]:
        load_in_4bit = training_args.load_in_kbits == 4
        load_in_8bit = training_args.load_in_kbits == 8
        if training_args.modules_to_save is not None:
            load_in_8bit_skip_modules = training_args.modules_to_save.split(',')
        else:
            load_in_8bit_skip_modules = None
        quantization_config = BitsAndBytesConfig(
            load_in_4bit=training_args.load_in_kbits == 4,
            load_in_8bit=training_args.load_in_kbits == 8,
            llm_int8_threshold=6.0,
            load_in_8bit_skip_modules=load_in_8bit_skip_modules,
            bnb_4bit_compute_dtype=compute_dtype,
            bnb_4bit_use_double_quant=training_args.double_quant,
            bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'}
        )
    else:
        load_in_4bit = False
        load_in_8bit = False
        quantization_config = None
    if quantization_config is not None:
        logger.info(f"quantization_config:{quantization_config.to_dict()}")
    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)
        )
        device_map = {"":int(os.environ.get("LOCAL_RANK") or 0)}
        logger.info(f"use_flash_attn:{training_args.flash_attn}")
        model = LlamaForCausalLM.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,
            use_auth_token=True if model_args.use_auth_token else None,
            torch_dtype=torch_dtype,
            low_cpu_mem_usage=True,
            device_map=device_map,
            load_in_4bit=load_in_4bit,
            load_in_8bit=load_in_8bit,
            use_flash_attention_2=training_args.flash_attn,
            quantization_config=quantization_config,
        )
    else:
        # TODO: flash_attn settng https://github.com/huggingface/transformers/issues/26878
        model = AutoModelForCausalLM.from_config(config)
        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")
    if training_args.load_in_kbits in [4, 8]:
        model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing)
    model.config.use_cache = False
    model_vocab_size = model.get_output_embeddings().weight.size(0)
    tokenizer_vocab_size = len(tokenizer)
    logger.info(f"Model vocab size: {model_vocab_size}")
    logger.info(f"Tokenizer vocab size: {tokenizer_vocab_size}")
    if tokenizer_vocab_size != jp_vocab_size:
        raise ValueError(f"The vocab size of tokenizer is {tokenizer_vocab_size}, not {jp_vocab_size}. Please use Japanese-LLaMA-2 tokenizer.")

    if model_vocab_size != tokenizer_vocab_size:
        new_nvocab_size = len(tokenizer)
        logger.info(f"Resize model vocab size to {new_nvocab_size}")

    # NOTE: Use at least recent version of transformers and deepspeed as of 2023/10,
    # otherwise resize may not take effect.
    #
    # https://github.com/huggingface/transformers/pull/26387 
    #
    # for efficiency, make internal embedding size multiple of 64(51470 => 51520)
    model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=64)
    print(model)

    if training_args.peft_path is not None:
        logger.info("Peft from pre-trained model")
        model = PeftModel.from_pretrained(model, training_args.peft_path, device_map=device_map)
    else:
        logger.info("Init new peft model")
        target_modules = training_args.trainable.split(',')
        modules_to_save = training_args.modules_to_save
        if modules_to_save is not None:
            modules_to_save = modules_to_save.split(',')
        lora_rank = training_args.lora_rank
        lora_dropout = training_args.lora_dropout
        lora_alpha = training_args.lora_alpha
        logger.info(f"target_modules: {target_modules}")
        logger.info(f"lora_rank: {lora_rank}")
        peft_config = LoraConfig(
            task_type=TaskType.CAUSAL_LM,
            target_modules=target_modules,
            inference_mode=False,
            r=lora_rank, lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            modules_to_save=modules_to_save)
        model = get_peft_model(model, peft_config)
    if training_args.gradient_checkpointing and \
        (not model.modules_to_save or 'embed_tokens' not in model.modules_to_save):
        # enable requires_grad to avoid exception during backward pass when using gradient_checkpoint without tuning embed.
        if hasattr(model.base_model, "enable_input_require_grads"):
            model.base_model.enable_input_require_grads()
        elif hasattr(model.base_model, "get_input_embeddings"):
            def make_inputs_require_grad(_module, _input, _output):
                _output.requires_grad_(True)
            model.base_model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
    for name, module in model.named_modules():
        if isinstance(module, LoraLayer):
            if training_args.bf16:
                module = module.to(torch.bfloat16)
            if training_args.fp16:
                module = module.to(torch.float16)
        if 'norm' in name:
            module = module.to(torch.float16)
        if 'lm_head' in name or 'embed_tokens' in name:
            if hasattr(module, 'weight'):
                if training_args.bf16 and module.weight.dtype == torch.float32:
                    module = module.to(torch.bfloat16)
                if training_args.fp16 and module.weight.dtype == torch.float32:
                    module = module.to(torch.float16)
    model.print_trainable_parameters()
    logger.info(f"model.modules_to_save: {model.modules_to_save}")

    # https://github.com/huggingface/peft/issues/286
    # updating state_dict fails to save LoRA weight on 2+ save_pretrained().
    # it looks we can simply uncomment it.
    #old_state_dict = model.state_dict
    #model.state_dict = (
    #    lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
    #).__get__(model, type(model))

    # 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,
        tokenizer=tokenizer,
        data_collator=fault_tolerance_data_collator,
        compute_metrics=compute_metrics if training_args.do_eval and not is_torch_tpu_available() else None,
        preprocess_logits_for_metrics=preprocess_logits_for_metrics
        if training_args.do_eval and not is_torch_tpu_available()
        else None,
    )
    trainer.add_callback(SavePeftModelCallback)
    # 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)

        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)


def test():
    japanese_tokenizer_dir = "../data/merged_tokenizer_hf"
    base_llama_model = "PY007/TinyLlama-1.1B-intermediate-step-480k-1T"


    japanese_tokenizer = LlamaTokenizer.from_pretrained(japanese_tokenizer_dir)
    print(japanese_tokenizer)

    assert len(japanese_tokenizer) == jp_vocab_size

    model = LlamaForCausalLM.from_pretrained(base_llama_model)
    model_vocab_size = model.get_output_embeddings().weight.size(0)
    print(model_vocab_size)
    model.resize_token_embeddings(len(japanese_tokenizer))

    #model.print_trainable_parameters()
    #print(model.p_vocab_size)

    # comma separated string
    trainable = "q_proj,v_proj"
    lora_rank = 8
    lora_dropout = 0.1
    lora_alpha = 32.0
    modules_to_save = None # or str

    target_modules = trainable.split(',')
    peft_config = LoraConfig(
        task_type=TaskType.CAUSAL_LM,
        target_modules=target_modules,
        r=lora_rank,
        inference_mode=False,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        modules_to_save=modules_to_save)

    model = get_peft_model(model, peft_config)
    model.print_trainable_parameters()

    #old_state_dict = model.state_dict

    ## ?
    #model.state_dict = (
    #    lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
    #).__get__(model, type(model))

if __name__ == "__main__":
    main()