muon-optimizer

Muon: An optimizer for the hidden layers of neural networks

This repo contains an implementation of the Muon optimizer originally described in this thread and this writeup.

Installation

pip install git+https://github.com/KellerJordan/Muon

Usage

Muon is intended to optimize only the internal ≥2D parameters of a network. Embeddings, classifier heads, and internal gains/biases should be optimized using AdamW.

# optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4, betas=(0.90, 0.95), weight_decay=0.01)

from muon import MuonWithAuxAdam
# Find ≥2D parameters in the body of the network -- these should be optimized by Muon
hidden_weights = [p for p in model.body.parameters() if p.ndim >= 2]
# Find everything else -- these should be optimized by AdamW
hidden_gains_biases = [p for p in model.body.parameters() if p.ndim < 2]
exterior_weights = [*model.head.parameters(), *model.embed.parameters()])
# Create the optimizer
# Note: you can also use multiple groups of each type with different hparams if you want.
muon_group = dict(params=hidden_weights, lr=0.02, weight_decay=0.01, use_muon=True)
adam_group = dict(params=hidden_gains_biases+exterior_weights, lr=3e-4,
                  betas=(0.9, 0.95), weight_decay=0.01, use_muon=False)
optimizer = MuonWithAuxAdam([muon_group, adam_group])

You'll have to replace model.body, model.head, and model.embed with whatever subset is appropriate for your model. E.g., for a ConvNet, Muon should optimize all the convolutional filters except the first one, and AdamW should optimize everything else.

Example usage

Example use in the NanoGPT speedrun

Example use in the CIFAR-10 speedrun

Hyperparameter tuning

Typically, the default values of momentum (0.95), nesterov (True), and ns_steps (5) work well. The only hyperparameter which must be tuned is the learning rate. It should have constant muP scaling, that is, as you scale up the model size, you shouldn't need to retune the learning rate.

Benchmarks

For a comparison between AdamW, Shampoo, SOAP, and Muon for training a 124M-parameter transformer, see here.

Accomplishments

• Lowered the record for training to 94% on CIFAR-10 from 3.3 A100-seconds to 2.6 A100-seconds
• Used to train a transformer to GPT-2 (XL) performance in $175 of compute
• Improved the training speed record for attaining GPT-2 (small) performance by a factor of 1.35x
• Used by the Kimi.ai frontier lab for scaled LLM training

More learning resources and results about Muon

• Blog post on Muon by Jialin Su (the creator of RoPE)
• Blog post by Jeremy Bernstein on theoretical background of Muon
• Tech report by Kimi.ai on using Muon for scaled training
• Why we chose Muon: Our chain of thought (by Jianlin Su at Kimi.ai)

Citation

@misc{jordan2024muon,
  author       = {Keller Jordan and Yuchen Jin and Vlado Boza and You Jiacheng and
                  Franz Cesista and Laker Newhouse and Jeremy Bernstein},
  title        = {Muon: An optimizer for hidden layers in neural networks},
  year         = {2024},
  url          = {https://kellerjordan.github.io/posts/muon/}
}