Function get_batch

Megatron-LM/evaluate_gpt2.py:142–185

get_batch subdivides the source data into chunks of length args.seq_length. If source is equal to the example output of the data loading example, with a seq_length limit of 2, we'd get the following two Variables for i = 0: ┌ a g m s ┐ ┌ b h n t ┐ └ b h n t ┘ └ c i o u ┘

(data_iterator, args, timers)

    return attention_mask, loss_mask, position_ids

def get_batch(data_iterator, args, timers):
    ''' get_batch subdivides the source data into chunks of
    length args.seq_length. If source is equal to the example
    output of the data loading example, with a seq_length limit
    of 2, we'd get the following two Variables for i = 0:
    ┌ a g m s ┐ ┌ b h n t ┐
    └ b h n t ┘ └ c i o u ┘
    Note that despite the name of the function, the subdivison of data is not
    done along the batch dimension (i.e. dimension 1), since that was handled
    by the data loader. The chunks are along dimension 0, corresponding
    to the seq_len dimension in the LSTM. A Variable representing an appropriate
    shard reset mask of the same dimensions is also returned.
    '''
    # Items and their type.
    keys = ['text', 'pad_mask']
    datatype = torch.int64

    # Broadcast data.
    timers('data loader').start()
    if data_iterator is not None:
        data = next(data_iterator)
    else:
        data = None
    timers('data loader').stop()
    data_b = mpu.broadcast_data(keys, data, datatype)

    # Unpack.
    tokens_ = data_b['text'].long()
    lm_labels = tokens_[:, 1:].contiguous()
    tokens = tokens_[:, :-1].contiguous()
    padding_mask = data_b['pad_mask'].byte()

    # Get the masks and postition ids.
    attention_mask, loss_mask, position_ids = get_masks_and_position_ids(
        tokens,
        args.eod_token,
        args.reset_position_ids,
        args.reset_attention_mask)

    # Convert
    if args.fp16:
        attention_mask = attention_mask.half()

    return tokens, lm_labels, attention_mask, position_ids, padding_mask


def forward_step(data_iterator, model, args, timers):