--moe-router-topk-scaling-factor argument works wrongly
Open
bug
community-request
module: moe
- Dominant language
- Python
- Stars
- 17.9k
- Forks
- 4.5k
- Avg merge
- 4d 6h
- Merged PRs (30d)
- 271
Description
Document here (https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/training/arguments.py#L3104) said that it only works when --moe-router-pre-softmax enabled, but it affects outputs whether --moe-router-pre-softmax option is on/off.
moe_router_topk_scaling_factor should become None if --moe-router-pre-softmax option is disabled, or documentation should be fixed.
Contributor guide
Assessment
This issue has not been assessed yet.