Struct MuonConfig 

pub struct MuonConfig { /* private fields */ }

Muon configuration.

Muon is an optimizer specifically designed for 2D parameters of neural network hidden layers (weight matrices). Other parameters such as biases and embeddings should be optimized using a standard method such as AdamW.

Learning Rate Adjustment

Muon adjusts the learning rate based on parameter shape to maintain consistent RMS across rectangular matrices. Two methods are available:

• Original: Uses sqrt(max(1, A/B)) where A and B are the first two dimensions. This is Keller Jordan’s method and is the default.

• MatchRmsAdamW: Uses 0.2 * sqrt(max(A, B)). This is Moonshot’s method designed to match AdamW’s RMS, allowing direct reuse of AdamW hyperparameters.

Example

use burn_optim::{MuonConfig, AdjustLrFn};

// Using default (Original) method
let optimizer = MuonConfig::new().init();

// Using MatchRmsAdamW for AdamW-compatible hyperparameters
let optimizer = MuonConfig::new()
    .with_adjust_lr_fn(AdjustLrFn::MatchRmsAdamW)
    .init();

References

• Muon: An optimizer for hidden layers in neural networks
• Muon is Scalable for LLM Training
• PyTorch Implementation
• Original Implementation

Implementations

impl MuonConfig

pub fn new() -> Self

Create a new instance of the config.

impl MuonConfig

pub fn with_momentum(self, momentum: MomentumConfig) -> Self

Momentum config.

pub fn with_ns_coefficients(self, ns_coefficients: (f32, f32, f32)) -> Self

Newton-Schulz iteration coefficients (a, b, c).

pub fn with_epsilon(self, epsilon: f32) -> Self

Epsilon for numerical stability.

pub fn with_ns_steps(self, ns_steps: usize) -> Self

Number of Newton-Schulz iteration steps.

pub fn with_adjust_lr_fn(self, adjust_lr_fn: AdjustLrFn) -> Self

Learning rate adjustment method.

pub fn with_weight_decay(self, weight_decay: Option<WeightDecayConfig>) -> Self

Set the default value for the field.

impl MuonConfig

pub fn init<B: AutodiffBackend, M: AutodiffModule<B>>( &self, ) -> OptimizerAdaptor<Muon<B::InnerBackend>, M, B>

Initialize Muon optimizer.

Returns

Returns an optimizer adaptor that can be used to optimize a module.

Example

use burn_optim::{MuonConfig, AdjustLrFn, decay::WeightDecayConfig};

// Basic configuration with default (Original) LR adjustment
let optimizer = MuonConfig::new()
    .with_weight_decay(Some(WeightDecayConfig::new(0.01)))
    .init();

// With AdamW-compatible settings using MatchRmsAdamW
let optimizer = MuonConfig::new()
    .with_adjust_lr_fn(AdjustLrFn::MatchRmsAdamW)
    .with_weight_decay(Some(WeightDecayConfig::new(0.1)))
    .init();

// Custom momentum and NS settings
let optimizer = MuonConfig::new()
    .with_momentum(MomentumConfig {
        momentum: 0.9,
        dampening: 0.1,
        nesterov: false,
    })
    .with_ns_steps(7)
    .init();

Trait Implementations

impl Clone for MuonConfig

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Config for MuonConfig

fn save<P>(&self, file: P) -> Result<(), Error>

where P: AsRef<Path>,

Available on crate feature std only.

Saves the configuration to a file.

fn load<P>(file: P) -> Result<Self, ConfigError>

where P: AsRef<Path>,

Available on crate feature std only.

Loads the configuration from a file.

fn load_binary(data: &[u8]) -> Result<Self, ConfigError>

Loads the configuration from a binary buffer.

impl Debug for MuonConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for MuonConfig

fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>

where D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Display for MuonConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Serialize for MuonConfig

fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>

where S: Serializer,

Serialize this value into the given Serde serializer.