[QUESTION] MoE score for aux loss
Open
community-request
module: moe
question
waiting-on-customer
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
I noticed that compute_routing_scores_for_aux_loss in megatron/core/transformer/moe/moe_utils.py does not handle grouped topk routing, while topk_routing_with_score_function does. This may cause the auxiliary loss to optimize for a different routing distribution than what is actually used for token dispatching.
Is this intentional behavior, or is it a bug?
If intentional, what is the reasoning behind using different routing logic for aux loss vs. actual token dispatching?
Contributor guide
Assessment
This issue has not been assessed yet.