megatron.utils.average_losses_across_data_parallel_group# megatron.utils.average_losses_across_data_parallel_group(losses)# Reduce a tensor of losses across all GPUs.