[bug] reduce_aux_losses_tracker_across_ranks all_reduce bug with num_layers==pp stage
Open
bug
community-request
module: moe
waiting-on-customer
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 3h
- Merged PRs (30d)
- 272
Description
in https://github.com/NVIDIA/Megatron-LM/blob/5153663876b322972d90115ee6b4c3894445c5ac/megatron/core/transformer/moe/moe_utils.py#L756
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
Contributor guide
Assessment
This issue has not been assessed yet.