NVIDIA / NVIDIA/Megatron-LM

Add MLA support for Muon optimizer

Open
#4,091 15 comments 2 reactions 1 assignee Claimed by @janEbert View on GitHub
community-request enhancement waiting-on-customer
Dominant language
Python
Stars
17.9k
Forks
4.5k
Avg merge
4d 6h
Merged PRs (30d)
271

Description

**Is your feature request related to a problem? Please describe.**
Currently, the Muon optimizer implementation in `megatron/core/optimizer/muon.py` lacks support for MLA (Multi-Latent Attention) parameters, as indicated by the TODO comment (`TODO(deyuf): support MLA`).

MLA-related parameters are not tagged with `is_qkv` and thus undergo direct orthogonalization. Maybe it lead to some unpredictable result.

**Describe the solution you'd like**
Add MLA support to the Muon optimizer.

Contributor guide

Open the contributing guide

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.