NVIDIA-NeMo / NVIDIA-NeMo/Automodel
rope_fusion broken with cp for MoEs
Open
@HuiyingLi is already working on this.
Since Aug 27, 2026.
PoR
- Dominant language
- Python
- Stars
- 963
- Forks
- 318
- Avg merge
- 3d 20h
- Merged PRs (30d)
- 143
Description
Enabling rope_fusion with cp > 1 gives incorrect loss for Qwen3 MoE 30b for long context training. Workaround for now is to disable rope_fusion for cp > 1, need to investigate root cause for proper fix.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.