[bug] reduce_aux_losses_tracker_across_ranks all_reduce bug with num_layers==pp stage

[lk137095576]

in

Megatron-LM/megatron/core/transformer/moe/moe_utils.py

Line 756 in 5153663

def reduce_aux_losses_tracker_across_ranks(track_names: Optional[List[str]] = None):

in reduce_aux_losses_tracker_across_ranks, if pp stage == model block，like num_layers=4, pp=4, all_reduce run error
torch.distributed.all_reduce(
values, group=parallel_state.get_pipeline_model_parallel_group()
)
if num_layers=4, pp=2, run right