Function loss_func

examples/linear_llama3/pretrain_llama.py:186–215

Loss function. Args: loss_mask (torch.Tensor): Used to mask out some portions of the loss output_tensor (torch.Tensor): The tensor with the losses

(loss_mask: torch.Tensor, output_tensor: torch.Tensor)

    return batch.values()

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

    Args:
        loss_mask (torch.Tensor): Used to mask out some portions of the loss
        output_tensor (torch.Tensor): The tensor with the losses
    """
    args = get_args()

    losses = output_tensor.float()
    loss_mask = loss_mask.view(-1).float()
    if args.context_parallel_size > 1:
        loss = torch.cat([torch.sum(losses.view(-1) * loss_mask).view(1), loss_mask.sum().view(1)])
        torch.distributed.all_reduce(loss, group=mpu.get_context_parallel_group())
        loss = loss[0] / loss[1]
    else:
        loss = torch.sum(losses.view(-1) * loss_mask) / loss_mask.sum()

    # Check individual rank losses are not NaN prior to DP all-reduce.
    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 * args.context_parallel_size, {'lm loss': averaged_loss[0]}


def forward_step(data_iterator, model: GPTModel):