Request for More Detailed Documentation on MuSGD Optimizer Usage

[Brayan532]

Hi Ultralytics team,

First of all, thank you for releasing YOLO26 and all the significant advancements you’ve brought to computer vision workflows.
I’ve read the official YOLO26 documentation and blog posts that highlight MuSGD as a hybrid optimizer combining SGD with Muon-style optimization to drive stable training. However, I couldn’t find concrete examples or recommended usage patterns (such as code snippets, parameter settings, or API usage) in the documentation that show how to enable and configure MuSGD in user workflows, especially for custom models. This information would be extremely helpful for the community.

Thanks very much for your continued investment in the Ultralytics ecosystem and for making cutting-edge CV technology more accessible!

Sincerely,
Brayan.

[UltralyticsAssistant]

👋 Hello @Brayan532, thank you for your interest in Ultralytics 🚀! This is an automated response to help you get unblocked quickly—an Ultralytics engineer will also take a look and assist soon. We recommend a visit to the Docs for new users where you can find many Python and CLI usage examples and where many of the most common questions may already be answered.

If this is a 🐛 Bug Report (for example, if MuSGD isn’t behaving as expected or you’re seeing errors enabling it), please provide a minimum reproducible example to help us debug it (including your command/script, a small config snippet, and the exact logs).

If this is a custom training ❓ Question (like recommended MuSGD usage patterns, configuration knobs, or best-practice settings for custom models), please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results ✅.

Join the Ultralytics community where it suits you best. For real-time chat, head to Discord 🎧. Prefer in-depth discussions? Check out Discourse. Or dive into threads on our Subreddit to share knowledge with the community.

Upgrade

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

[glenn-jocher]

Hi @Brayan532 — you can enable MuSGD in standard Ultralytics training by setting optimizer="MuSGD" (or leaving optimizer="auto", which will select MuSGD for longer runs based on total iterations); for the exact behavior (which params get Muon vs plain SGD, and the internal Muon/SGD blend used), see the MuSGD/Muon API reference and the trainer’s optimizer construction logic.

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=100, imgsz=640, optimizer="MuSGD")

If you’re training a fully custom PyTorch loop (not model.train()), you can also import MuSGD directly from ultralytics.optim.muon and pass explicit param groups (with use_muon=True only on 2D+ tensors), as shown in the same MuSGD/Muon API reference; feel free to reply here with your model type (detect/seg/pose) and whether you’re fine-tuning or training from scratch if you want a recommended starting lr0/momentum for your setup.

[Brayan532]

Dear Glenn,
I'm grateful for your swift and clear reply.

Best wishes,
B.

[glenn-jocher]

Thanks Brayan—glad that clarified it, and credit to the broader YOLO community and the Ultralytics team; if you want feedback on MuSGD settings for your specific dataset/model (fine-tune vs scratch), just reply here with your model.train(...) args and a short snippet of your train logs (and if anything errors, please include the full traceback plus yolo checks output).