qlora.py

# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

from collections import defaultdict
import copy
import json
import os
from os.path import exists, join, isdir
from dataclasses import dataclass, field
import sys
from typing import Optional, Dict, Sequence
import numpy as np
from tqdm import tqdm
import logging
import bitsandbytes as bnb

import torch
import transformers
from torch.nn.utils.rnn import pad_sequence
import argparse
from transformers import (
    AutoTokenizer, 
    AutoModelForCausalLM, 
    set_seed, 
    Seq2SeqTrainer,
    BitsAndBytesConfig,
    LlamaTokenizerFast

)
from datasets import load_dataset
import evaluate

from peft import (
    prepare_model_for_int8_training,
    LoraConfig,
    get_peft_model,
    get_peft_model_state_dict,
    PeftModel
)
from peft.tuners.lora import LoraLayer
from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR


torch.backends.cuda.matmul.allow_tf32 = True

logger = logging.getLogger(__name__)

IGNORE_INDEX = -100
DEFAULT_PAD_TOKEN = "[PAD]"

@dataclass
class ModelArguments:
    model_name_or_path: Optional[str] = field(
        default="EleutherAI/pythia-12b"
    )
    trust_remote_code: Optional[bool] = field(
        default=False,
        metadata={"help": "Enable unpickling of arbitrary code in AutoModelForCausalLM#from_pretrained."}
    )

@dataclass
class DataArguments:
    eval_dataset_size: int = field(
        default=1024, metadata={"help": "Size of validation dataset."}
    )
    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."
        },
    )
    source_max_len: int = field(
        default=1024,
        metadata={"help": "Maximum source sequence length. Sequences will be right padded (and possibly truncated)."},
    )
    target_max_len: int = field(
        default=256,
        metadata={"help": "Maximum target sequence length. Sequences will be right padded (and possibly truncated)."},
    )
    dataset: str = field(
        default='alpaca',
        metadata={"help": "Which dataset to finetune on. See datamodule for options."}
    )

@dataclass
class TrainingArguments(transformers.Seq2SeqTrainingArguments):
    cache_dir: Optional[str] = field(
        default=None
    )
    train_on_source: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to train on the input in addition to the target text."}
    )
    mmlu_split: Optional[str] = field(
        default='eval',
        metadata={"help": "The MMLU split to run on"}
    )
    mmlu_dataset: Optional[str] = field(
        default='mmlu-fs',
        metadata={"help": "MMLU dataset to use: options are `mmlu-zs` for zero-shot or `mmlu-fs` for few shot."}
    )
    do_mmlu_eval: Optional[bool] = field(
        default=False,
        metadata={"help": "Whether to run the MMLU evaluation."}
    )
    max_mmlu_samples: Optional[int] = field(
        default=None,
        metadata={"help": "If set, only evaluates on `max_mmlu_samples` of the MMMLU dataset."}
    )
    mmlu_source_max_len: int = field(
        default=2048,
        metadata={"help": "Maximum source sequence length for mmlu."}
    )
    full_finetune: bool = field(
        default=False,
        metadata={"help": "Finetune the entire model without adapters."}
    )
    adam8bit: bool = field(
        default=False,
        metadata={"help": "Use 8-bit adam."}
    )
    double_quant: bool = field(
        default=True,
        metadata={"help": "Compress the quantization statistics through double quantization."}
    )
    quant_type: str = field(
        default="nf4",
        metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
    )
    bits: int = field(
        default=4,
        metadata={"help": "How many bits to use."}
    )
    lora_r: int = field(
        default=64,
        metadata={"help": "Lora R dimension."}
    )
    lora_alpha: float = field(
        default=16,
        metadata={"help": " Lora alpha."}
    )
    lora_dropout: float = field(
        default=0.0,
        metadata={"help":"Lora dropout."}
    )
    max_memory_MB: int = field(
        default=80000,
        metadata={"help": "Free memory per gpu."}
    )
    report_to: str = field(
        default='none',
        metadata={"help": "To use wandb or something else for reporting."}
    )
    output_dir: str = field(default='./output', metadata={"help": 'The output dir for logs and checkpoints'})
    optim: str = field(default='paged_adamw_32bit', metadata={"help": 'The optimizer to be used'})
    per_device_train_batch_size: int = field(default=1, metadata={"help": 'The training batch size per GPU. Increase for better speed.'})
    gradient_accumulation_steps: int = field(default=16, metadata={"help": 'How many gradients to accumulate before to perform an optimizer step'})
    max_steps: int = field(default=10000, metadata={"help": 'How many optimizer update steps to take'})
    weight_decay: float = field(default=0.0, metadata={"help": 'The L2 weight decay rate of AdamW'}) # use lora dropout instead for regularization if needed
    learning_rate: float = field(default=0.0002, metadata={"help": 'The learnign rate'})
    remove_unused_columns: bool = field(default=False, metadata={"help": 'Removed unused columns. Needed to make this codebase work.'})
    max_grad_norm: float = field(default=0.3, metadata={"help": 'Gradient clipping max norm. This is tuned and works well for all models tested.'})
    gradient_checkpointing: bool = field(default=True, metadata={"help": 'Use gradient checkpointing. You want to use this.'})
    do_train: bool = field(default=True, metadata={"help": 'To train or not to train, that is the question?'})
    lr_scheduler_type: str = field(default='constant', metadata={"help": 'Learning rate schedule. Constant a bit better than cosine, and has advantage for analysis'})
    warmup_ratio: float = field(default=0.03, metadata={"help": 'Fraction of steps to do a warmup for'})
    logging_steps: int = field(default=10, metadata={"help": 'The frequency of update steps after which to log the loss'})
    group_by_length: bool = field(default=True, metadata={"help": 'Group sequences into batches with same length. Saves memory and speeds up training considerably.'})
    save_strategy: str = field(default='steps', metadata={"help": 'When to save checkpoints'})
    save_steps: int = field(default=250, metadata={"help": 'How often to save a model'})
    save_total_limit: int = field(default=40, metadata={"help": 'How many checkpoints to save before the oldest is overwritten'})

@dataclass
class GenerationArguments:
    # For more hyperparameters check:
    # https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationConfig
    # Length arguments
    max_new_tokens: Optional[int] = field(
        default=256,
        metadata={"help": "Maximum number of new tokens to be generated in evaluation or prediction loops"
                          "if predict_with_generate is set."}
    )
    min_new_tokens : Optional[int] = field(
        default=None,
        metadata={"help": "Minimum number of new tokens to generate."}
    )

    # Generation strategy
    do_sample: Optional[bool] = field(default=False)
    num_beams: Optional[int] = field(default=1)
    num_beam_groups: Optional[int] = field(default=1)
    penalty_alpha: Optional[float] = field(default=None)
    use_cache: Optional[bool] = field(default=True) 

    # Hyperparameters for logit manipulation
    temperature: Optional[float] = field(default=1.0)
    top_k: Optional[int] = field(default=50)
    top_p: Optional[float] = field(default=1.0)
    typical_p: Optional[float] = field(default=1.0)
    diversity_penalty: Optional[float] = field(default=0.0) 
    repetition_penalty: Optional[float] = field(default=1.0) 
    length_penalty: Optional[float] = field(default=1.0)
    no_repeat_ngram_size: Optional[int] = field(default=0) 

def find_all_linear_names(args, model):
    cls = bnb.nn.Linear4bit if args.bits == 4 else (bnb.nn.Linear8bitLt if args.bits == 8 else torch.nn.Linear)
    lora_module_names = set()
    for name, module in model.named_modules():
        if isinstance(module, cls):
            names = name.split('.')
            lora_module_names.add(names[0] if len(names) == 1 else names[-1])


    if 'lm_head' in lora_module_names: # needed for 16-bit
        lora_module_names.remove('lm_head')
    return list(lora_module_names)


class SavePeftModelCallback(transformers.TrainerCallback):
    def save_model(self, args, state, kwargs):
        print('Saving PEFT checkpoint...')
        if state.best_model_checkpoint is not None:
            checkpoint_folder = os.path.join(state.best_model_checkpoint, "adapter_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, "adapter_model")
        kwargs["model"].save_pretrained(peft_model_path)

        pytorch_model_path = os.path.join(checkpoint_folder, "pytorch_model.bin")
        if os.path.exists(pytorch_model_path):
            os.remove(pytorch_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):
        def touch(fname, times=None):
            with open(fname, 'a'):
                os.utime(fname, times)

        touch(join(args.output_dir, 'completed'))
        self.save_model(args, state, kwargs)

def get_accelerate_model(args, checkpoint_dir):

    n_gpus = torch.cuda.device_count()
    max_memory = f'{args.max_memory_MB}MB'
    max_memory = {i: max_memory for i in range(n_gpus)}

    if args.full_finetune: assert args.bits in [16, 32]

    print(f'loading base model {args.model_name_or_path}...')
    compute_dtype = (torch.float16 if args.fp16 else (torch.bfloat16 if args.bf16 else torch.float32))
    model = AutoModelForCausalLM.from_pretrained(
        args.model_name_or_path,
        load_in_4bit=args.bits == 4,
        load_in_8bit=args.bits == 8,
        device_map='auto',
        max_memory=max_memory,
        quantization_config=BitsAndBytesConfig(
            load_in_4bit=args.bits == 4,
            load_in_8bit=args.bits == 8,
            llm_int8_threshold=6.0,
            llm_int8_has_fp16_weight=False,
            bnb_4bit_compute_dtype=compute_dtype,
            bnb_4bit_use_double_quant=args.double_quant,
            bnb_4bit_quant_type=args.quant_type # {'fp4', 'nf4'}
        ),
        torch_dtype=(torch.float32 if args.fp16 else (torch.bfloat16 if args.bf16 else torch.float32)),
        trust_remote_code=args.trust_remote_code,
    )
    if compute_dtype == torch.float16 and args.bits == 4:
        major, minor = torch.cuda.get_device_capability()
        if major >= 8:
            print('='*80)
            print('Your GPU supports bfloat16, you can accelerate training with the argument --bf16')
            print('='*80)

    setattr(model, 'model_parallel', True)
    setattr(model, 'is_parallelizable', True)
    modules = find_all_linear_names(args, model)

    model.config.torch_dtype=(torch.float32 if args.fp16 else (torch.bfloat16 if args.bf16 else torch.float32))

    if not args.full_finetune:
        model = prepare_model_for_int8_training(model, use_gradient_checkpointing=args.gradient_checkpointing)
    if args.gradient_checkpointing:
        model.gradient_checkpointing_enable()

    config = LoraConfig(
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        target_modules=modules,
        lora_dropout=args.lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
    )
    if not args.full_finetune:
        if checkpoint_dir is not None:
            print("Loading adapters from checkpoint.")
            model = PeftModel.from_pretrained(model, join(checkpoint_dir, 'adapter_model'))
            for name, p in model.named_parameters():
                if 'lora' in name:
                    print(name, p.sum())
        else:
            print(f'adding LoRA modules...')
            model = get_peft_model(model, config)

    if args.gradient_checkpointing:
        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)


    for name, module in model.named_modules():
        if isinstance(module, LoraLayer):
            if args.bf16:
                module = module.to(torch.bfloat16)
        if 'norm' in name:
            module = module.to(torch.float32)
        if 'lm_head' in name or 'embed_tokens' in name:
            if hasattr(module, 'weight'):
                if args.bf16 and module.weight.dtype == torch.float32:
                    module = module.to(torch.bfloat16)
    return model

def print_trainable_parameters(args, model):
    """
    Prints the number of trainable parameters in the model.
    """
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    if args.bits == 4: trainable_params /= 2
    print(f"trainable params: {trainable_params} || all params: {all_param} || trainable: {100 * trainable_params / all_param}")

def smart_tokenizer_and_embedding_resize(
    special_tokens_dict: Dict,
    tokenizer: transformers.PreTrainedTokenizer,
    model: transformers.PreTrainedModel,
):
    """Resize tokenizer and embedding.

    Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
    """
    num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
    model.resize_token_embeddings(len(tokenizer))

    if num_new_tokens > 0:
        input_embeddings = model.get_input_embeddings().weight.data
        output_embeddings = model.get_output_embeddings().weight.data

        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)

        input_embeddings[-num_new_tokens:] = input_embeddings_avg
        output_embeddings[-num_new_tokens:] = output_embeddings_avg

@dataclass
class DataCollatorForCausalLM(object):
    tokenizer: transformers.PreTrainedTokenizer
    source_max_len: int
    target_max_len: int
    train_on_source: bool
    predict_with_generate: bool

    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
        # Extract elements
        sources = [example['input'] for example in instances]
        targets = [f"{example['output']}{self.tokenizer.eos_token}" for example in instances]
        # Tokenize
        tokenized_sources_with_prompt = self.tokenizer(
            sources,
            max_length=self.source_max_len,
            truncation=True,
        )
        tokenized_targets = self.tokenizer(
            targets,
            max_length=self.target_max_len,
            truncation=True,
            add_special_tokens=False,
        )
        # Build the input and labels for causal LM
        input_ids = []
        labels = [] 
        for tokenized_source, tokenized_target in zip(
            tokenized_sources_with_prompt['input_ids'], 
            tokenized_targets['input_ids']
        ):
            if not self.predict_with_generate:
                input_ids.append(torch.tensor(tokenized_source + tokenized_target))
                if not self.train_on_source:
                    labels.append(
                        torch.tensor([IGNORE_INDEX for _ in range(len(tokenized_source))] + copy.deepcopy(tokenized_target))
                    )
                else:
                    labels.append(torch.tensor(copy.deepcopy(tokenized_source + tokenized_target)))
            else:
                input_ids.append(torch.tensor(tokenized_source))
        # Apply padding
        input_ids = pad_sequence(input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id)
        labels = pad_sequence(labels, batch_first=True, padding_value=IGNORE_INDEX) if not self.predict_with_generate else None
        data_dict = {
            'input_ids': input_ids,
            'attention_mask':input_ids.ne(self.tokenizer.pad_token_id),
        }
        if labels is not None:
            data_dict['labels'] = labels
        return data_dict

def extract_unnatural_instructions_data(examples, extract_reformulations=False):
    out = {
        'input': [],
        'output': [],
    }
    for example_instances in examples['instances']:
        for instance in example_instances:
            out['input'].append(instance['instruction_with_input'])
            out['output'].append(instance['output'])
    if extract_reformulations:
        for example_reformulations in examples['reformulations']:
            if example_reformulations is not None:
                for instance in example_reformulations:
                    out['input'].append(instance['instruction_with_input'])
                    out['output'].append(instance['output'])
    return out

PROMPT_DICT = {
    "prompt_input": (
        "Below is an instruction that describes a task, paired with an input that provides further context. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response: "
    ),
    "prompt_no_input": (
        "Below is an instruction that describes a task. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Response: "
    ),
}

def extract_alpaca_dataset(example):
    if example.get("input", "") != "":
        prompt_format = PROMPT_DICT["prompt_input"]
    else:
        prompt_format = PROMPT_DICT["prompt_no_input"]
    return {'input': prompt_format.format(**example)}

def make_data_module(tokenizer: transformers.PreTrainedTokenizer, args) -> Dict:
    """
    Make dataset and collator for supervised fine-tuning.
    Datasets are expected to have the following columns: { `input`, `output` }

    Available datasets to be selected with `dataset` argument:
        - alpaca, 52002 examples
        - alpaca cleaned, 51942 examples   
        - chip2 (OIG), 210289 examples
        - self-instruct, 82612 examples
        - hh-rlhf (Anthropic), 160800 examples
        - longform, 23.7k examples

    Coming soon:
        - unnatural instructions core, 66010 examples
        - unnatural instructions full, 240670 examples
        - alpaca-gpt4, 52002 examples
        - unnatural-instructions-gpt4, 9000 examples
        - oa-rlhf (OpenAssistant) primary message tree only, 9209 examples
        - oa-rlhf-assistant (OpenAssistant) all assistant  replies with ranking
        - supernatural-instructions, 69624 examples (same as paper with 100 ex/task more can be used)
        - flan (FLAN v2), up to 20M examples available

    Not Available:
        - vicuna, not released at the moment.
    """
    # Load dataset.
    # Alpaca
    if args.dataset == 'alpaca':
        dataset = load_dataset("tatsu-lab/alpaca")
        dataset = dataset.map(extract_alpaca_dataset, remove_columns=['instruction'])
    # Alpaca clean
    elif args.dataset == 'alpaca-clean':
        dataset = load_dataset("yahma/alpaca-cleaned")
        dataset = dataset.map(extract_alpaca_dataset, remove_columns=['instruction'])
    # Chip2
    elif args.dataset == 'chip2':
        dataset = load_dataset("laion/OIG", data_files='unified_chip2.jsonl')
        dataset = dataset.map(lambda x: {
            'input': x['text'].split('\n<bot>: ')[0].replace('<human>: ', ''),
            'output': x['text'].split('\n<bot>: ')[1],
        }, remove_columns=['text', 'metadata'])
    # Self Instruct
    elif args.dataset == 'self-instruct':
        dataset = load_dataset("yizhongw/self_instruct", name='self_instruct')
        for old, new in [["prompt", "input"], ["completion", "output"]]:
            dataset = dataset.rename_column(old, new)
    # Anthropic rlhf
    elif args.dataset == 'hh-rlhf':
        dataset = load_dataset("Anthropic/hh-rlhf")
        dataset = dataset.map(lambda x: {
            'input': '',
            'output': x['chosen']
        }, remove_columns=['chosen', 'rejected'])
    # LongForm
    elif args.dataset == 'longform':
        dataset = load_dataset("akoksal/LongForm")

    elif args.dataset == 'guanaco':
        dataset = load_dataset("timdettmers/openassistant-guanaco")
        dataset = dataset.map(lambda x: {
            'input': '',
            'output': x['text']
        }, remove_columns=['text'])

    elif args.dataset == 'vicuna':
        raise NotImplementedError("Vicuna data was not released.")
    else:
        raise NotImplementedError(f"Dataset {args.dataset} not implemented yet.")

    # Split train/eval, reduce size
    if args.do_eval or args.do_predict:
        if 'eval' in dataset:
            eval_dataset = dataset['eval']
        else:
            print('Splitting train dataset in train and validation according to `eval_dataset_size`')
            dataset = dataset["train"].train_test_split(
                test_size=args.eval_dataset_size, shuffle=True, seed=42
            )
            eval_dataset = dataset['test']
        if args.max_eval_samples is not None and len(eval_dataset) > args.max_eval_samples:
            eval_dataset = eval_dataset.select(range(args.max_eval_samples))
        if args.group_by_length:
            eval_dataset = eval_dataset.map(lambda x: {'length': len(x['input']) + len(x['output'])})
    if args.do_train:
        train_dataset = dataset['train']
        if args.max_train_samples is not None and len(train_dataset) > args.max_train_samples:
            train_dataset = train_dataset.select(range(args.max_train_samples))
        if args.group_by_length:
            train_dataset = train_dataset.map(lambda x: {'length': len(x['input']) + len(x['output'])})

    data_collator = DataCollatorForCausalLM(
        tokenizer=tokenizer, 
        source_max_len=args.source_max_len,
        target_max_len=args.target_max_len,
        train_on_source=args.train_on_source,
        predict_with_generate=args.predict_with_generate,
    )
    return dict(
        train_dataset=train_dataset if args.do_train else None, 
        eval_dataset=eval_dataset if args.do_eval else None,
        predict_dataset=eval_dataset if args.do_predict else None,
        data_collator=data_collator
    )

def get_last_checkpoint(checkpoint_dir):
    if isdir(checkpoint_dir):
        is_completed = exists(join(checkpoint_dir, 'completed'))
        if is_completed: return None, True # already finished
        max_step = 0
        for filename in os.listdir(checkpoint_dir):
            if isdir(join(checkpoint_dir, filename)) and filename.startswith('checkpoint'):
                max_step = max(max_step, int(filename.replace('checkpoint-', '')))
        if max_step == 0: return None, is_completed # training started, but no checkpoint
        checkpoint_dir = join(checkpoint_dir, f'checkpoint-{max_step}')
        print(f"Found a previous checkpoint at: {checkpoint_dir}")
        return checkpoint_dir, is_completed # checkpoint found!
    return None, False # first training

def train():
    hfparser = transformers.HfArgumentParser((
        ModelArguments, DataArguments, TrainingArguments, GenerationArguments
    ))
    model_args, data_args, training_args, generation_args, extra_args = \
        hfparser.parse_args_into_dataclasses(return_remaining_strings=True)
    training_args.generation_config = transformers.GenerationConfig(**vars(generation_args))
    args = argparse.Namespace(
        **vars(model_args), **vars(data_args), **vars(training_args)
    )
    

    checkpoint_dir, completed_training = get_last_checkpoint(args.output_dir)
    if completed_training:
        print('Detected that training was already completed!')

    model = get_accelerate_model(args, checkpoint_dir)
    training_args.skip_loading_checkpoint_weights=True

    model.config.use_cache = False
    print_trainable_parameters(args, model)
    print('loaded model')
    set_seed(args.seed)

    # Tokenizer
    tokenizer = AutoTokenizer.from_pretrained(
        args.model_name_or_path,
        cache_dir=args.cache_dir,
        padding_side="right",
        use_fast=True,
    )
    if tokenizer.pad_token is None:
        smart_tokenizer_and_embedding_resize(
            special_tokens_dict=dict(pad_token=DEFAULT_PAD_TOKEN),
            tokenizer=tokenizer,
            model=model,
        )
    if isinstance(tokenizer, LlamaTokenizerFast):
        # LLaMA tokenizer may not have correct special tokens set.
        # Check and add them if missing to prevent them from being parsed into different tokens.
        # Note that these are present in the vocabulary. 
        # Note also that `model.config.pad_token_id` is 0 which corresponds to `<unk>` token.
        if tokenizer.eos_token_id != model.config.eos_token_id or tokenizer.pad_token_id != model.config.pad_token_id or tokenizer.unk_token_id != model.config.unk_token_id:
            tokenizer.add_special_tokens(
                {
                    "eos_token": tokenizer.convert_ids_to_tokens(model.config.eos_token_id),
                    "bos_token": tokenizer.convert_ids_to_tokens(model.config.bos_token_id),
                    "unk_token": tokenizer.convert_ids_to_tokens(model.config.pad_token_id),
                }
            )

    data_module = make_data_module(tokenizer=tokenizer, args=args)
    trainer = Seq2SeqTrainer(
        model=model, 
        tokenizer=tokenizer,
        args=training_args,
        **{k:v for k,v in data_module.items() if k != 'predict_dataset'},
    )

    # Callbacks
    if not args.full_finetune:
        trainer.add_callback(SavePeftModelCallback)
    if args.do_mmlu_eval:
        if args.mmlu_dataset == 'mmlu-zs':
            mmlu_dataset = load_dataset("json", data_files={
                'eval': 'data/mmlu/zero_shot_mmlu_val.json',
                'test': 'data/mmlu/zero_shot_mmlu_test.json',
            })
            mmlu_dataset = mmlu_dataset.remove_columns('subject')
        # MMLU Five-shot (Eval/Test only)
        elif args.mmlu_dataset == 'mmlu' or args.mmlu_dataset == 'mmlu-fs':
            mmlu_dataset = load_dataset("json", data_files={
                'eval': 'data/mmlu/five_shot_mmlu_val.json',
                'test': 'data/mmlu/five_shot_mmlu_test.json',
            })
            # mmlu_dataset = mmlu_dataset.remove_columns('subject')
        mmlu_dataset = mmlu_dataset[args.mmlu_split]
        if args.max_mmlu_samples is not None:
            mmlu_dataset = mmlu_dataset.select(range(args.max_mmlu_samples))
        abcd_idx = [
            tokenizer("A", add_special_tokens=False).input_ids[0],
            tokenizer("B", add_special_tokens=False).input_ids[0],
            tokenizer("C", add_special_tokens=False).input_ids[0],
            tokenizer("D", add_special_tokens=False).input_ids[0],
        ]
        accuracy = evaluate.load("accuracy")

        class MMLUEvalCallback(transformers.TrainerCallback):
            def on_evaluate(self, args, state, control, model, **kwargs):
                data_loader = trainer.get_eval_dataloader(mmlu_dataset)
                source_max_len = trainer.data_collator.source_max_len
                trainer.data_collator.source_max_len = args.mmlu_source_max_len
                trainer.model.eval()
                preds, refs = [], []
                loss_mmlu = 0
                for batch in tqdm(data_loader, total=len(data_loader)):
                    (loss, logits, labels) = trainer.prediction_step(trainer.model,batch,prediction_loss_only=False,)
                    # There are two tokens, the output, and eos token.
                    for i, logit in enumerate(logits):
                        label_non_zero_id = (batch['labels'][i] != -100).nonzero()[0][0]
                        logit_abcd = logit[label_non_zero_id-1][abcd_idx]
                        preds.append(torch.argmax(logit_abcd).item())
                    labels = labels[labels != IGNORE_INDEX].view(-1, 2)[:,0]
                    refs += [abcd_idx.index(label) for label in labels.tolist()]
                    
                    loss_mmlu += loss.item()
                # Extract results by subject.
                results = {'mmlu_loss':loss_mmlu/len(data_loader)}
                subject = mmlu_dataset['subject']
                subjects = {s:{'refs':[], 'preds':[]} for s in set(subject)}
                for s,p,r in zip(subject, preds, refs):
                    subjects[s]['preds'].append(p)
                    subjects[s]['refs'].append(r)
                subject_scores = []
                for subject in subjects:
                    subject_score = accuracy.compute(
                        references=subjects[subject]['refs'],
                        predictions=subjects[subject]['preds']
                    )['accuracy']
                    results[f'mmlu_{args.mmlu_split}_accuracy_{subject}'] = subject_score
                    subject_scores.append(subject_score)
                results[f'mmlu_{args.mmlu_split}_accuracy'] = np.mean(subject_scores)
                trainer.log(results)
                trainer.data_collator.source_max_len = source_max_len

        trainer.add_callback(MMLUEvalCallback)

    # Verifying the datatypes.
    dtypes = {}
    for _, p in model.named_parameters():
        dtype = p.dtype
        if dtype not in dtypes: dtypes[dtype] = 0
        dtypes[dtype] += p.numel()
    total = 0
    for k, v in dtypes.items(): total+= v
    for k, v in dtypes.items():
        print(k, v, v/total)

    # if args.bits < 16:
    #     old_state_dict = model.state_dict
    #     model.state_dict = (
    #         lambda self, *_, **__: get_peft_model_state_dict(self, old_state_dict())
    #     ).__get__(model, type(model))

    all_metrics = {"run_name": args.run_name}
    # Training
    if args.do_train:
        train_result = trainer.train(resume_from_checkpoint=checkpoint_dir)
        metrics = train_result.metrics
        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()
        all_metrics.update(metrics)
    # Evaluation
    if args.do_eval:
        logger.info("*** Evaluate ***")
        metrics = trainer.evaluate(metric_key_prefix="eval")
        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)
        all_metrics.update(metrics)
    # Prediction
    if args.do_predict:
        logger.info("*** Predict ***")
        prediction_output = trainer.predict(test_dataset=data_module['predict_dataset'],metric_key_prefix="predict")
        prediction_metrics = prediction_output.metrics
        predictions = prediction_output.predictions
        predictions = np.where(predictions != -100, predictions, tokenizer.pad_token_id)
        predictions = tokenizer.batch_decode(
            predictions, skip_special_tokens=True, clean_up_tokenization_spaces=True
        )
        with open(os.path.join(args.output_dir, 'predictions.jsonl'), 'w') as fout:
            for i, example in enumerate(data_module['predict_dataset']):
                example['prediction_with_input'] = predictions[i].strip()
                example['prediction'] = predictions[i].replace(example['input'], '').strip()
                fout.write(json.dumps(example) + '\n')
        print(prediction_metrics)
        trainer.log_metrics("predict", prediction_metrics)
        trainer.save_metrics("predict", prediction_metrics)
        all_metrics.update(prediction_metrics)

    if (args.do_train or args.do_eval or args.do_predict):
        with open(os.path.join(args.output_dir, "metrics.json"), "w") as fout:
            fout.write(json.dumps(all_metrics))

if __name__ == "__main__":
    train()