pretrain_gpt.py

# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
"""Pretrain GPT."""

import os
import torch
from torch import Tensor
from functools import partial
from typing import Union
from megatron import get_args
from megatron import print_rank_0
from megatron import get_timers
from megatron import get_tokenizer
from megatron.core import tensor_parallel
from megatron.core.enums import ModelType
from megatron.data.gpt_dataset import GPTDataset, build_train_valid_test_datasets
import megatron.model
from megatron.training import pretrain
from megatron.core.transformer.spec_utils import import_module
from megatron.utils import get_ltor_masks_and_position_ids
from megatron.utils import average_losses_across_data_parallel_group
from megatron.arguments import core_transformer_config_from_args


def model_provider(pre_process=True, post_process=True) -> megatron.model.GPTModel:
    """Builds the model.

    If you set the use_mcore_models to True, it will return the mcore GPT model and if not the legacy GPT model.

    Args:
        pre_process (bool, optional): Set to true if you need to compute embedings. Defaults to True.
        post_process (bool, optional): Set to true if you need to want to compute output logits/loss. Defaults to True.


    Returns:
        Union[GPTModel, megatron.model.GPTModel]: The returned model
    """

    print_rank_0('building GPT model ...')
    config = core_transformer_config_from_args(get_args())

    model = megatron.model.GPTModel(
        config,
        num_tokentypes=0,
        parallel_output=True,
        pre_process=pre_process,
        post_process=post_process
    )

    return model


def get_batch(data_iterator):
    """Generate a batch."""
    args = get_args()
    tokenizer = get_tokenizer()

    # Items and their type.
    keys = ['text']
    datatype = torch.int64

    # Broadcast data.
    if data_iterator is not None:
        data = next(data_iterator)
    else:
        data = None
    data_b = tensor_parallel.broadcast_data(keys, data, datatype)

    # Unpack.
    tokens_ = data_b['text'].long()
    labels = tokens_[:, 1:].contiguous()
    tokens = tokens_[:, :-1].contiguous()

    # Get the masks and postition ids.
    attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
        tokens,
        tokenizer.eod,
        args.reset_position_ids,
        args.reset_attention_mask,
        args.eod_mask_loss)

    return tokens, labels, loss_mask, attention_mask, position_ids

def loss_func(loss_mask: Tensor, output_tensor: Tensor):
    """Loss function.

    Args:
        loss_mask (Tensor): Used to mask out some portions of the loss
        output_tensor (Tensor): The tensor with the losses
    """    
    losses = output_tensor.float()
    loss_mask = loss_mask.view(-1).float()
    loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()

    # Check individual rank losses are not NaN prior to DP all-reduce.
    args = get_args()
    if args.check_for_nan_in_loss_and_grad:
        global_rank = torch.distributed.get_rank()
        assert not loss.isnan(), (
            f'Rank {global_rank}: found NaN in local forward loss calculation. '
            f'Device: {torch.cuda.current_device()}, node: {os.uname()[1]}'
        )

    # Reduce loss for logging.
    averaged_loss = average_losses_across_data_parallel_group([loss])

    return loss, {'lm loss': averaged_loss[0]}


def forward_step(data_iterator, model):
    """Forward training step.

    Args:
        data_iterator : Input data iterator
        model (GPTModel): The GPT Model
    """
    args = get_args()
    timers = get_timers()

    # Get the batch.
    timers('batch-generator', log_level=2).start()
    tokens, labels, loss_mask, attention_mask, position_ids = get_batch(
        data_iterator)
    timers('batch-generator').stop()

    output_tensor = model(tokens, position_ids, attention_mask,
                          labels=labels)

    return output_tensor, partial(loss_func, loss_mask)


def train_valid_test_datasets_provider(train_val_test_num_samples):
    """Build the train test and validation datasets.

    Args:
        train_val_test_num_samples : A list containing the number of samples in train test and validation.
    """
    args = get_args()

    print_rank_0('> building train, validation, and test datasets '
                 'for GPT ...')
    train_ds, valid_ds, test_ds = build_train_valid_test_datasets(
        data_prefix=args.data_path,
        splits_string=args.split,
        train_valid_test_num_samples=train_val_test_num_samples,
        seq_length=args.seq_length,
        seed=args.seed,
        skip_warmup=(not args.mmap_warmup),
        train_data_prefix=args.train_data_path,
        valid_data_prefix=args.valid_data_path,
        test_data_prefix=args.test_data_path,
        data_cache_path=args.data_cache_path)
    print_rank_0("> finished creating GPT datasets ...")

    return train_ds, valid_ds, test_ds


if __name__ == "__main__":

    pretrain(train_valid_test_datasets_provider,
             model_provider,
             ModelType.encoder_or_decoder,
             forward_step,
             args_defaults={'tokenizer_type': 'GPT2BPETokenizer'})