train_sparse.py

# SPDX-License-Identifier: Apache-2.0
import copy
import os
import sys
import warnings
from typing import Any, Dict, List, Optional, Union

import torch
from composer import Trainer
from composer.core import Algorithm, Evaluator, Event
from composer.utils import dist, get_device, reproducibility
from omegaconf import DictConfig, ListConfig
from omegaconf import OmegaConf as om
from transformers import PreTrainedTokenizerBase

from llmfoundry import (COMPOSER_MODEL_REGISTRY, ComposerHFCausalLM,
                        MPTForCausalLM, build_finetuning_dataloader,
                        build_text_denoising_dataloader)
from llmfoundry.data.text_data import build_text_dataloader
from llmfoundry.utils.builders import (build_algorithm, build_callback,
                                       build_icl_evaluators, build_logger,
                                       build_optimizer, build_scheduler,
                                       build_tokenizer)
from llmfoundry.utils.config_utils import (log_config, pop_config,
                                           process_init_device,
                                           update_batch_size_info)

import torch.nn.functional as TF
from composer.trainer.dist_strategy import prepare_fsdp_module
from composer.utils import get_device


class MaskPrunedWeights(Algorithm):
    def match(self, event, state):
        # masking weights after optimizer step should be sufficient
        # if we detect weird behaviour, we can also do it before the forward pass
        # by adding `or event == Event.BATCH_START`
        return event == Event.BATCH_END

    @torch.no_grad()
    def apply(self, event, state, logger):
        def mask_weights(module):
            if hasattr(module, 'mask'):
                module.weight *= module.mask

        state.model.apply(mask_weights)


def kldiv_loss(student_logits, teacher_logits, temperature):
    "Kullback-Leibler divergence loss"
    num_tokens = student_logits.numel() / student_logits.size(-1)
    return (
        TF.kl_div(
            input=TF.log_softmax(student_logits / temperature, dim=-1),
            target=TF.log_softmax(teacher_logits / temperature, dim=-1),
            log_target=True,
            reduction="sum",
        )
        * (temperature**2)
        / num_tokens
    )


class KnowledgeDistillation(Algorithm):
    def __init__(self, teacher, temperature, hardness_ce, hardness_kldiv, hardness_squarehead):
        self.teacher = teacher
        self.temperature = temperature
        # loss = hardness_ce x CrossEntropyLoss + hardness_kldiv x KLDivLoss + hardness_squarehead x SquareHeadLoss
        self.hardness_ce = hardness_ce
        self.hardness_kldiv = hardness_kldiv
        self.hardness_squarehead = hardness_squarehead
        self.first_time = True

    def match(self, event, state):
        """
        Event.AFTER_LOSS = augment loss with knowledge distillation
        Event.FIT_START = initialize FSDP for teacher model
        Event.BEFORE_FORWARD = enable hidden states for SquareHead KD
        """
        return event == Event.AFTER_LOSS or event == Event.FIT_START or event == Event.BEFORE_FORWARD

    def apply(self, event, state, logger):
        if event == Event.FIT_START and self.first_time:
            self.first_time = False  # just to be sure FIT_START is reached only once
            if torch.distributed.get_world_size() > 1:
                prepare_fsdp_module(
                    model=self.teacher,
                    optimizers=None,
                    fsdp_config=state.fsdp_config,
                    precision=state.precision,
                    device=get_device(None),
                    auto_microbatching=False,
                )
            else:
                self.teacher = self.teacher.to("cuda")
        elif event == Event.BEFORE_FORWARD:
            state.batch["output_hidden_states"] = True
        elif event == Event.AFTER_LOSS:
            with torch.no_grad():
                teacher_outputs = self.teacher(state.batch)

            loss_gen_tokens = state.batch['labels'] != -100
            student_logits = state.outputs.logits[loss_gen_tokens]
            teacher_logits = teacher_outputs.logits[loss_gen_tokens]

            kl_loss = self.hardness_kldiv * kldiv_loss(student_logits, teacher_logits, self.temperature)
            squarehead_loss = torch.tensor(0.0)

            if self.hardness_squarehead > 0:
                layerwise_losses = []
                for i in range(1, len(state.outputs.hidden_states)):
                    useful_tokens = state.batch['attention_mask'] == 1
                    student_states = state.outputs.hidden_states[i][useful_tokens]
                    teacher_states = teacher_outputs.hidden_states[i][useful_tokens]
                    layerwise_losses.append((student_states - teacher_states).pow(2).mean() / (teacher_states.pow(2).mean() + torch.finfo(torch.bfloat16).eps))

                squarehead_loss = self.hardness_squarehead * sum(layerwise_losses)

            to_log = {
                "losses/ce": self.hardness_ce * state.loss.item(),
                "losses/kldiv": kl_loss.item(),
                "losses/squarehead": squarehead_loss.item(),
            }
            logger.log_metrics(to_log)

            state.loss *= self.hardness_ce
            state.loss += kl_loss + squarehead_loss


def validate_config(cfg: DictConfig):
    """Validates compatible model and dataloader selection."""
    loaders = [cfg.train_loader]
    if 'eval_loader' in cfg:
        loaders.append(cfg.eval_loader)
    for loader in loaders:
        if loader.name == 'text':
            if cfg.model.name in ['hf_prefix_lm', 'hf_t5']:
                raise ValueError(
                    f'Model type "{cfg.model.name}" is not supported when using the "text " ' +\
                    f'dataloader. Please use the "text_denoising" dataloader to pre-train that model type.')
        elif loader.name == 'text_denoising':
            if cfg.model.name == 'hf_causal_lm':
                raise ValueError(
                    f'Model type "{cfg.model.name}" is not supported when using the "text_denoising" ' +\
                    f'dataloader. Please use the "text" dataloader to pre-train that model type.')
            if loader.mixture_of_denoisers.decoder_only_format and cfg.model.name == 'hf_t5':
                warnings.warn(
                    'Model type "hf_t5" requires `decoder_only_format` to be ``False``. ' +\
                    'Overriding `decoder_only_format` from ``True`` to ``False``.')
                loader.mixture_of_denoisers.decoder_only_format = False
            if (not loader.mixture_of_denoisers.decoder_only_format
               ) and cfg.model.name == 'hf_prefix_lm':
                warnings.warn(
                    'Model type "hf_prefix_lm" requires `decoder_only_format` to be ``True``. ' +\
                    'Overriding `decoder_only_format` from ``False`` to ``True``.')
                loader.mixture_of_denoisers.decoder_only_format = True

    if 'icl_tasks' in cfg:
        if cfg.model.name == 'hf_t5':
            raise ValueError(
                'ICL evaluation does not currently support Encoder-Decoder models, such as "hf_t5".'
            )

    if (cfg.model.get('fc_type', 'torch') != 'te' and 'te' not in cfg.model.get(
            'ffn_config', {}).get('ffn_type', 'mptmlp') and
            'fp8' in cfg.precision):
        warnings.warn(
            "fp8 only supported for te.Linear layers. Either set `cfg.model.fc_typ='te'` or "
            +
            "`cfg.model.ffn_config.ffn_type='te_ln_mlp'` to enable layers using fp8 precision."
        )

    if (cfg.model.get('fc_type', 'torch') == 'te' or
            'te' in cfg.model.get('ffn_config', {}).get('ffn_type', 'mptmlp')):
        fsdp_config = cfg.get('fsdp_config', None)
        act_ckpt = fsdp_config.get('activation_checkpointing', False)
        act_ckpt_reentrant = fsdp_config.get(
            'activation_checkpointing_reentrant', True)
        if fsdp_config is not None and act_ckpt == True and act_ckpt_reentrant == False:
            warnings.warn(
                '`te.Linear` layers do not support activation_checkpointing with '
                + '`activation_checkpointing_reentrant = False`. ' +
                'Setting cfg.fsdp_config.activation_checkpointing_reentrant=True.'
            )
            cfg.fsdp_config.activation_checkpointing_reentrant = True

    if 'te' in cfg.model.get('ffn_config', {}).get('ffn_type', 'mptmlp'):
        warnings.warn(
            '`te.LayerNormMLP` requires has issues with torch._dynamo. ' +
            'Setting `torch._dynamo.config.suppress_errors = True` and falling back to eager.'
        )
        torch._dynamo.config.suppress_errors = True  # type: ignore


def build_composer_model(model_cfg: DictConfig,
                         tokenizer: PreTrainedTokenizerBase):
    warnings.filterwarnings(
        action='ignore',
        message='Torchmetrics v0.9 introduced a new argument class property')
    if model_cfg.name not in COMPOSER_MODEL_REGISTRY:
        raise ValueError(
            f'Not sure how to build model with name={model_cfg.name}')
    return COMPOSER_MODEL_REGISTRY[model_cfg.name](model_cfg, tokenizer)


def build_composer_peft_model(
        pretrained_model_name_or_path: str, lora_args: Dict[str, Any],
        tokenizer: PreTrainedTokenizerBase) -> ComposerHFCausalLM:
    try:
        from peft import LoraConfig, get_peft_model
    except ImportError as e:
        raise ImportError(
            'Error importing from peft. Please verify that peft and peft utils '
            +
            'are installed by running `pip install -e .[peft]` from `llm-foundry/`. '
            + f'Error encountered: {e}')

    # 1) loads a hf model, 2) adds peft modules, 3) wraps it in a ComposerHFCausalLM.
    print('Building Lora config...')
    lora_cfg = LoraConfig(**lora_args)

    print('Building model from HuggingFace checkpoint...')
    model = MPTForCausalLM.from_pretrained(pretrained_model_name_or_path,
                                           trust_remote_code=True)
    print('Model built!')

    print('Adding Lora modules...')
    model = get_peft_model(model, lora_cfg)
    print('Lora modules added!')

    model = ComposerHFCausalLM(model, tokenizer)

    return model


def print_trainable_parameters(model: torch.nn.Module) -> None:
    # 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()
    print(
        f'trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}'
    )


def build_dataloader(cfg: DictConfig, tokenizer: PreTrainedTokenizerBase,
                     device_batch_size: int):
    if cfg.name == 'text':
        return build_text_dataloader(
            cfg,
            tokenizer,
            device_batch_size,
        )
    elif cfg.name == 'text_denoising':
        return build_text_denoising_dataloader(
            cfg,
            tokenizer,
            device_batch_size,
        )
    elif cfg.name == 'finetuning':
        return build_finetuning_dataloader(
            cfg,
            tokenizer,
            device_batch_size,
        )
    else:
        raise ValueError(f'Not sure how to build dataloader with config: {cfg}')


def main(cfg: DictConfig):
    # Filter deprecation warning from torch internal usage
    warnings.filterwarnings(
        action='ignore',
        category=UserWarning,
        message=
        'torch.distributed.*_base is a private function and will be deprecated.*'
    )

    # Check for incompatibilities between the model and data loaders
    validate_config(cfg)

    # Resolve all interpolation variables as early as possible
    om.resolve(cfg)

    # Create copy of config for logging
    logged_cfg: DictConfig = copy.deepcopy(cfg)

    # Get max split size mb
    max_split_size_mb: Optional[int] = cfg.pop('max_split_size_mb', None)
    if max_split_size_mb is not None:
        os.environ[
            'PYTORCH_CUDA_ALLOC_CONF'] = f'max_split_size_mb:{max_split_size_mb}'

    # Set seed first
    seed: int = pop_config(cfg, 'seed', must_exist=True)
    reproducibility.seed_all(seed)

    # Initialize pytorch distributed training process groups
    dist_timeout: Union[int, float] = pop_config(cfg,
                                                 'dist_timeout',
                                                 must_exist=False,
                                                 default_value=600.0)
    dist.initialize_dist(get_device(None), timeout=dist_timeout)

    # Get global and device batch size information from distributed/single node setting
    cfg = update_batch_size_info(cfg)
    logged_cfg.update(cfg, merge=True)

    # Mandatory model training configs
    model_config: DictConfig = pop_config(cfg, 'model', must_exist=True)
    tokenizer_config: DictConfig = pop_config(cfg, 'tokenizer', must_exist=True)
    optimizer_config: Dict[str, Any] = pop_config(cfg,
                                                  'optimizer',
                                                  must_exist=True,
                                                  convert=True)
    scheduler_config: Dict[str, Any] = pop_config(cfg,
                                                  'scheduler',
                                                  must_exist=True,
                                                  convert=True)
    train_loader_config: DictConfig = pop_config(cfg,
                                                 'train_loader',
                                                 must_exist=True)

    # Optional fsdp data, fine-tuning, and eval configs
    fsdp_config: Optional[Dict[str, Any]] = pop_config(cfg,
                                                       'fsdp_config',
                                                       must_exist=False,
                                                       default_value=None,
                                                       convert=True)
    lora_config: Optional[Dict[str, Any]] = pop_config(cfg,
                                                       'lora',
                                                       must_exist=False,
                                                       default_value=None,
                                                       convert=True)
    eval_loader_config: Optional[DictConfig] = pop_config(cfg,
                                                          'eval_loader',
                                                          must_exist=False,
                                                          default_value=None)
    icl_tasks_config: Optional[ListConfig] = pop_config(cfg,
                                                        'icl_tasks',
                                                        must_exist=False,
                                                        default_value=None)

    # Optional logging, evaluation and callback configs
    logger_configs: Optional[DictConfig] = pop_config(cfg,
                                                      'loggers',
                                                      must_exist=False,
                                                      default_value=None)
    callback_configs: Optional[DictConfig] = pop_config(cfg,
                                                        'callbacks',
                                                        must_exist=False,
                                                        default_value=None)
    algorithm_configs: Optional[DictConfig] = pop_config(cfg,
                                                         'algorithms',
                                                         must_exist=False,
                                                         default_value=None)

    # Mandatory hyperparameters for training
    device_train_batch_size: int = pop_config(cfg,
                                              'device_train_batch_size',
                                              must_exist=True)
    device_eval_batch_size: int = pop_config(cfg,
                                             'device_eval_batch_size',
                                             must_exist=True)
    max_duration: Union[int, str] = pop_config(cfg,
                                               'max_duration',
                                               must_exist=True)
    eval_interval: Union[int, str] = pop_config(cfg,
                                                'eval_interval',
                                                must_exist=True)
    precision: str = pop_config(cfg, 'precision', must_exist=True)
    max_seq_len: int = pop_config(cfg, 'max_seq_len', must_exist=True)

    # Optional parameters will be set to default values if not specified.
    default_run_name: str = os.environ.get('RUN_NAME', 'llm')
    run_name: str = pop_config(cfg,
                               'run_name',
                               must_exist=False,
                               default_value=default_run_name)
    save_folder: Optional[str] = pop_config(cfg,
                                            'save_folder',
                                            must_exist=False,
                                            default_value=None)
    save_latest_filename: str = pop_config(cfg,
                                           'save_latest_filename',
                                           must_exist=False,
                                           default_value='latest-rank{rank}.pt')
    save_overwrite: bool = pop_config(cfg,
                                      'save_overwrite',
                                      must_exist=False,
                                      default_value=False)
    save_weights_only: bool = pop_config(cfg,
                                         'save_weights_only',
                                         must_exist=False,
                                         default_value=False)
    save_filename: str = pop_config(
        cfg,
        'save_filename',
        must_exist=False,
        default_value='ep{epoch}-ba{batch}-rank{rank}.pt')
    save_interval: Union[str, int] = pop_config(cfg,
                                                'save_interval',
                                                must_exist=False,
                                                default_value='1000ba')
    save_num_checkpoints_to_keep: int = pop_config(
        cfg, 'save_num_checkpoints_to_keep', must_exist=False, default_value=-1)
    progress_bar = pop_config(cfg,
                              'progress_bar',
                              must_exist=False,
                              default_value=False)
    log_to_console: bool = pop_config(cfg,
                                      'log_to_console',
                                      must_exist=False,
                                      default_value=True)
    python_log_level: str = pop_config(cfg,
                                       'python_log_level',
                                       must_exist=False,
                                       default_value='debug')
    console_log_interval: Union[int, str] = pop_config(cfg,
                                                       'console_log_interval',
                                                       must_exist=False,
                                                       default_value='1ba')
    device_train_microbatch_size: Union[str, int] = pop_config(
        cfg,
        'device_train_microbatch_size',
        must_exist=False,
        default_value='auto')
    eval_subset_num_batches: int = pop_config(cfg,
                                              'eval_subset_num_batches',
                                              must_exist=False,
                                              default_value=-1)
    eval_first: bool = pop_config(cfg,
                                  'eval_first',
                                  must_exist=False,
                                  default_value=False)
    load_path: str = pop_config(cfg,
                                'load_path',
                                must_exist=False,
                                default_value=None)
    load_weights_only: bool = pop_config(cfg,
                                         'load_weights_only',
                                         must_exist=False,
                                         default_value=False)
    load_ignore_keys: Optional[List[str]] = pop_config(cfg,
                                                       'load_ignore_keys',
                                                       must_exist=False,
                                                       default_value=None)
    # Enable autoresume from model checkpoints if possible
    autoresume_default: bool = False
    if logged_cfg.get('run_name', None) is not None \
        and save_folder is not None \
        and not save_overwrite \
        and not save_weights_only:
        print('As run_name, save_folder, and save_latest_filename are set, \
                changing autoresume default to True...')
        autoresume_default = True
    autoresume: bool = pop_config(cfg,
                                  'autoresume',
                                  must_exist=False,
                                  default_value=autoresume_default)

    # Pop known unused parameters that are used as interpolation variables or
    # created by update_batch_size_info.
    pop_config(cfg, 'data_local', must_exist=False)
    pop_config(cfg, 'data_remote', must_exist=False)
    pop_config(cfg, 'global_seed', must_exist=False)
    pop_config(cfg, 'global_train_batch_size', must_exist=False)
    pop_config(cfg, 'n_gpus', must_exist=False)
    pop_config(cfg, 'device_train_grad_accum', must_exist=False)

    # Warn users for unused parameters
    for key in cfg:
        warnings.warn(
            f'Unused parameter {key} found in cfg. Please check your yaml to ensure this parameter is necessary.'
        )

    # Warn if fsdp is enabled but user only has 1 GPU
    if dist.get_world_size() == 1 and fsdp_config is not None:
        warnings.warn(
            'FSDP is not applicable for single-GPU training. Reverting to DDP.')
        fsdp_config = None

    # Initialize context
    init_context = process_init_device(model_config, fsdp_config)
    logged_cfg.update({'fsdp_config': fsdp_config}, merge=True)

    # Build tokenizer
    tokenizer = build_tokenizer(tokenizer_config)

    # Build Model
    print('Initializing model...')
    with init_context:
        if lora_config is not None:  # frozen model + trainable lora modules
            model: ComposerHFCausalLM = build_composer_peft_model(
                model_config.pretrained_model_name_or_path, lora_config['args'],
                tokenizer)
            print_trainable_parameters(model)  # should not be 100%
        else:  # standard model
            model = build_composer_model(model_config, tokenizer)
            if "knowledge_distillation" in cfg and cfg.knowledge_distillation.teacher_name_or_path is not None:
                print(f"[Debugging] Knowledge Distillation config = {cfg.knowledge_distillation}")
                teacher_config = copy.deepcopy(model_config)
                teacher_config['pretrained_model_name_or_path'] = cfg.knowledge_distillation.teacher_name_or_path
                teacher = build_composer_model(teacher_config, tokenizer)
                teacher.eval()

    def attach_masks(model, to_layer):
        for name, module in model.named_children():
            if isinstance(module, torch.nn.Linear) and torch.sum(module.weight == 0)/module.weight.numel() >= 0.1:
                mask = torch.where(module.weight == 0, torch.tensor(0, dtype=torch.uint8), torch.tensor(1, dtype=torch.uint8))
                module.register_buffer("mask", mask, persistent=False)
                print(f"[Debugging] attaching mask to {name} with sparsity = {torch.sum(mask == 0)/mask.numel()}")
            else:
                attach_masks(module, to_layer)

    attach_masks(model, torch.nn.Linear)

    # Log number of parameters
    n_params = sum(p.numel() for p in model.parameters())
    logged_cfg.update({'n_params': n_params})

    # Optimizer
    optimizer_name: str = optimizer_config.pop('name')
    optimizer = build_optimizer(model, optimizer_name, optimizer_config)

    # Scheduler
    scheduler_name: str = scheduler_config.pop('name')
    scheduler = build_scheduler(scheduler_name, scheduler_config)

    # Loggers
    loggers = [
        build_logger(str(name), logger_cfg)
        for name, logger_cfg in logger_configs.items()
    ] if logger_configs else None

    # Callbacks
    callbacks = [
        build_callback(str(name), callback_cfg)
        for name, callback_cfg in callback_configs.items()
    ] if callback_configs else None

    # Algorithms
    algorithms = [
        build_algorithm(str(name), algorithm_cfg)
        for name, algorithm_cfg in algorithm_configs.items()
    ] if algorithm_configs else None

    if algorithms is None:
        algorithms = [MaskPrunedWeights()]
    else:
        algorithms.append(MaskPrunedWeights())

    if "knowledge_distillation" in cfg and cfg.knowledge_distillation.teacher_name_or_path is not None:
        algorithms.append(KnowledgeDistillation(teacher, cfg.knowledge_distillation.temperature, cfg.knowledge_distillation.hardness_ce, cfg.knowledge_distillation.hardness_kldiv, cfg.knowledge_distillation.hardness_squarehead))

    # Dataloaders
    print('Building train loader...')
    train_loader = build_dataloader(
        train_loader_config,
        tokenizer,
        device_train_batch_size,
    )

    ## Evaluation
    print('Building eval loader...')
    evaluators = []
    if eval_loader_config is not None:
        assert model.train_metrics is not None
        eval_dataloader = build_dataloader(eval_loader_config, tokenizer,
                                           device_eval_batch_size)
        eval_metric_names = list(model.train_metrics.keys())
        eval_loader = Evaluator(label='eval',
                                dataloader=eval_dataloader,
                                metric_names=eval_metric_names)
        evaluators.append(eval_loader)

    if icl_tasks_config is not None:
        icl_evaluators, _ = build_icl_evaluators(icl_tasks_config, tokenizer,
                                                 max_seq_len,
                                                 device_eval_batch_size)
        evaluators.extend(icl_evaluators)

    # Build the Trainer
    print('Building trainer...')
    trainer = Trainer(
        run_name=run_name,
        seed=seed,
        model=model,
        train_dataloader=train_loader,
        eval_dataloader=evaluators,
        optimizers=optimizer,
        schedulers=scheduler,
        max_duration=max_duration,
        eval_interval=eval_interval,
        eval_subset_num_batches=eval_subset_num_batches,
        progress_bar=progress_bar,
        log_to_console=log_to_console,
        console_log_interval=console_log_interval,
        loggers=loggers,
        callbacks=callbacks,
        precision=precision,
        algorithms=algorithms,
        device_train_microbatch_size=device_train_microbatch_size,
        fsdp_config=fsdp_config,  # type: ignore
        save_folder=save_folder,
        save_filename=save_filename,
        save_latest_filename=save_latest_filename,
        save_interval=save_interval,
        save_num_checkpoints_to_keep=save_num_checkpoints_to_keep,
        save_overwrite=save_overwrite,
        save_weights_only=save_weights_only,
        load_path=load_path,
        load_weights_only=load_weights_only,
        load_ignore_keys=load_ignore_keys,
        autoresume=autoresume,
        python_log_level=python_log_level,
        dist_timeout=dist_timeout,
    )

    print('Logging config')
    log_config(logged_cfg)
    torch.cuda.empty_cache()

    # Eval first if requested
    if eval_first and trainer.state.timestamp.batch.value == 0:
        trainer.eval()

    print('Starting training...')
    trainer.fit()
    print('Done.')

    print('Saving directly into HF-friendly format')
    if "WANDB_PROJECT" in os.environ and os.environ["WANDB_DISABLED"] == "False":
        path_to_save = os.path.join("output_dir", os.environ["WANDB_PROJECT"], run_name)
    else:
        path_to_save = os.path.join("output_dir", run_name)

    if torch.distributed.get_rank() == 0:
        os.makedirs(path_to_save, exist_ok=True) # <-- override if it exists
        # manually copy over the source code so that we dont depend on HF/transformers nor Mosaic/llmfoundry implementations of the MPT model
        if os.path.exists(model_config.pretrained_model_name_or_path):
            import shutil
            files_to_copy = [
                "adapt_tokenizer.py",
                "attention.py",
                "blocks.py",
                "config.json",
                "configuration_mpt.py",
                "custom_embedding.py",
                "flash_attn_triton.py",
                "generation_config.json",
                "hf_prefixlm_converter.py",
                "meta_init_context.py",
                "modeling_mpt.py",
                "norm.py",
                "param_init_fns.py",
                "special_tokens_map.json",
                "tokenizer.json",
                "tokenizer_config.json",
            ]
            print(f"[Debugging] Manually copying source code for MPT from {model_config.pretrained_model_name_or_path} to {path_to_save}")
            for f in files_to_copy:
                print(f"[Debugging] Copying {f}...")
                shutil.copyfile(os.path.join(cfg.model_name_or_path, f), os.path.join(path_to_save, f))

    torch.distributed.barrier()
    # Save the model in sharded format
    from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, StateDictType, FullStateDictConfig
    full_state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
    with FSDP.state_dict_type(model.model, StateDictType.FULL_STATE_DICT, full_state_dict_config):
        model.model.save_pretrained(path_to_save, is_main_process=torch.distributed.get_rank() == 0, state_dict=model.model.state_dict())

    # NOTE: for some reason the saving code above would create empty pytorch_model.bin file, so we delete it manually
    # TODO: figure out why this happens
    if torch.distributed.get_rank() == 0 and os.path.exists(os.path.join(path_to_save, "pytorch_model.bin")):
        tmp = torch.load(os.path.join(path_to_save, "pytorch_model.bin"))
        if not tmp:  # empty dict, remove it
            os.remove(os.path.join(path_to_save, "pytorch_model.bin"))


if __name__ == '__main__':
    yaml_path, args_list = sys.argv[1], sys.argv[2:]
    with open(yaml_path) as f:
        yaml_cfg = om.load(f)
    cli_cfg = om.from_cli(args_list)
    cfg = om.merge(yaml_cfg, cli_cfg)
    assert isinstance(cfg, DictConfig)
    main(cfg)