linkedin / linkedin/Liger-Kernel
Support GMPO Chunked loss
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 6.6k
- Forks
- 603
- Avg merge
- 1d 20h
- Merged PRs (30d)
- 47
Description
Please add native support for Geometric-Mean Policy Optimization (GMPO) to Liger’s chunked GRPO loss.
GMPO is a policy optimization algorithm published at ICLR 2026 that improves training stability by applying token-level clipping and aggregating importance ratios with a sequence-level geometric mean. Using DeepSeek-R1-Distill-Qwen-7B, GMPO improves the average Pass@1 score across five mathematical reasoning benchmarks from GRPO’s 59.3 to 63.4.
official code:https://github.com/callsys/GMPO
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.
Research direction
Start by reading Liger’s existing chunked GRPO loss implementation, then compare its behavior with the official GMPO code linked in the issue. The work is complete when native GMPO support is integrated into that loss and its token-level clipping and sequence-level geometric-mean aggregation are covered by the project’s relevant validation.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Mostly clear
- Newbie friendliness
- 45/100