[QUESTION] Why not gather routing_map in sequence_load_balancing_loss_func function
Open
community-request
module: moe
question
waiting-on-maintainers
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
**Why not gather routing_map in sequence_load_balancing_loss_func function **
Here is the code at https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/transformer/moe/moe_utils.py#L117
Here do a gather for 'probs_for_aux_loss' but not for 'routing_map'. Why? Shouldn't it be the same?
Contributor guide
Assessment
This issue has not been assessed yet.