tf.keras.optimizers.Muon

Optimizer that implements the Muon algorithm.

Inherits From: Optimizer

tf.keras.optimizers.Muon(
    learning_rate=0.001,
    adam_beta_1=0.9,
    adam_beta_2=0.999,
    adam_weight_decay=0.004,
    epsilon=1e-07,
    weight_decay=0.004,
    clipnorm=None,
    clipvalue=None,
    global_clipnorm=None,
    use_ema=False,
    ema_momentum=0.99,
    ema_overwrite_frequency=None,
    loss_scale_factor=None,
    gradient_accumulation_steps=None,
    name=None,
    exclude_layers=None,
    exclude_embeddings=True,
    muon_a=3.4445,
    muon_b=-4.775,
    muon_c=2.0315,
    adam_lr_ratio=1,
    momentum=0.95,
    ns_steps=5,
    nesterov=True,
    rms_rate=0.2,
    **kwargs
)

Note that this optimizer should not be used in the following layers:

  1. Embedding layer
  2. Final output fully connected layer
  3. Any {0,1}-D variables

These should all be optimized using AdamW.

The Muon optimizer can use both the Muon update step or the AdamW update step based on the following:

  • For any variable that isn't 2D, the AdamW step will be used. This is not configurable.
  • If the argument exclude_embeddings (defaults to True) is set to True, the AdamW step will be used.
  • For any variablewith a name that matches an expression listed in the argument exclude_layers (a list), the AdamW step will be used.
  • Any other variable uses the Muon step.

Typically, you only need to pass the name of your densely-connected output layer to exclude_layers, e.g. exclude_layers=["output_dense"].

learning_rate A float, keras.optimizers.schedules.LearningRateSchedule instance, or a callable that takes no arguments and returns the actual value to use. The learning rate. Defaults to 0.001.
adam_beta_1 A float value or a constant float tensor, or a callable that takes no arguments and returns the actual value to use. The exponential decay rate for the 1st moment estimates. Defaults to 0.9.
adam_beta_2 A float value or a constant float tensor, or a callable that takes no arguments and returns the actual value to use. The exponential decay rate for the 2nd moment estimates. Defaults to 0.999.
adam_weight_decay Float. If set, weight decay is applied when using the Adam optimizer.
epsilon A small constant for numerical stability. This is "epsilon hat" in the Kingma and Ba paper (in the formula just before Section 2.1), not the epsilon in Algorithm 1 of the paper. It be used at Adamw.Defaults to 1e-7.
exclude_layers List of strings, keywords of layer names to exclude. All layers with keywords in their path will use adamw.
exclude_embeddings Boolean value If True, embedding layers will use adamw.
muon_a Float, parameter a of the muon algorithm. It is recommended to use the default value
muon_b Float, parameter b of the muon algorithm. It is recommended to use the default value
muon_c Float, parameter c of the muon algorithm. It is recommended to use the default value
adam_lr_ratio Float, the ratio of the learning rate when using Adam to the main learning rate. It is recommended to set it to 1
momentum Float, momentum used by internal SGD.
ns_steps Integer, number of Newton-Schulz iterations to run.
nesterov Boolean, whether to use Nesterov-style momentum { {base_optimizer_keyword_args} }
rms_rate Float. A parameter from https://arxiv.org/abs/2502.16982 that can enhance the stability of Muon, allowing it to use the same learning rate and weight decay as Adam. Defaults to 0.2. Set to None to disable this feature.

iterations

learning_rate

variables

Methods

add_optimizer_variables

View source

add_optimizer_variables(
    trainable_variables, name, initializer='zeros'
)

Add optimizer variables from the list of trainable model variables.

Create an optimizer variable based on the information of the supplied model variables. For example, in SGD optimizer momemtum, for each model variable, a corresponding momemtum variable is created of the same shape and dtype.

Note that trainable variables with v.overwrite_with_gradient == True will insert None, into the output list, since the optimizer variable will not be used anyways, and could be wasteful.

Args
trainable_variables keras.Variable, the corresponding model variable to the optimizer variable to be created.
name The name prefix(es) of the optimizer variable(s) to be created. Can be a single string or list of strings. If a list of strings, will create an optimizer variable for each prefix. The variable name will follow the pattern {variable_name}_{trainable_variable.name}, e.g., momemtum/dense_1.
initializer Initializer object(s) to use to populate the initial variable value(s), or string name of a built-in initializer (e.g. "random_normal"). If unspecified, defaults to "zeros".

Returns
A list of optimizer variables, in the format of keras.Variables. If multiple names are provide, returns a tuple of lists.

add_variable

View source

add_variable(
    shape,
    initializer='zeros',
    dtype=None,
    aggregation='none',
    layout=None,
    name=None
)

Add a variable to the optimizer.

Args
shape Shape tuple for the variable. Must be fully-defined (no None entries).
initializer Initializer object to use to populate the initial variable value, or string name of a built-in initializer (e.g. "random_normal"). Defaults to "zeros".
dtype Dtype of the variable to create, e.g. "float32". If unspecified, defaults to the keras.backend.floatx().
aggregation Optional string, one of None, "none", "mean", "sum" or "only_first_replica". Annotates the variable with the type of multi-replica aggregation to be used for this variable when writing custom data parallel training loops. Defaults to "none".
layout Optional tensor layout. Defaults to None.
name String name of the variable. Useful for debugging purposes.

Returns
An optimizer variable, in the format of keras.Variable.

add_variable_from_reference

View source

add_variable_from_reference(
    reference_variable, name=None, initializer='zeros'
)

Add an optimizer variable from the model variable.

Create an optimizer variable based on the information of model variable. For example, in SGD optimizer momemtum, for each model variable, a corresponding momemtum variable is created of the same shape and dtype.

Args
reference_variable keras.Variable. The corresponding model variable to the optimizer variable to be created.
name Optional string. The name prefix of the optimizer variable to be created. If not provided, it will be set to "var". The variable name will follow the pattern {variable_name}_{reference_variable.name}, e.g., momemtum/dense_1. Defaults to None.
initializer Initializer object to use to populate the initial variable value, or string name of a built-in initializer (e.g. "random_normal"). If unspecified, defaults to "zeros".

Returns
An optimizer variable, in the format of keras.Variable.

apply

View source

apply(
    grads, trainable_variables=None
)

Update traininable variables according to provided gradient values.

grads should be a list of gradient tensors with 1:1 mapping to the list of variables the optimizer was built with.

trainable_variables can be provided on the first call to build the optimizer.

apply_gradients

View source

apply_gradients(
    grads_and_vars
)

assign

View source

assign(
    variable, value
)

Assign a value to a variable.

This should be used in optimizers instead of variable.assign(value) to support backend specific optimizations. Note that the variable can be a model variable or an optimizer variable; it can be a backend native variable or a Keras variable.

Args
variable The variable to update.
value The value to add to the variable.

assign_add

View source

assign_add(
    variable, value
)

Add a value to a variable.

This should be used in optimizers instead of variable.assign_add(value) to support backend specific optimizations. Note that the variable can be a model variable or an optimizer variable; it can be a backend native variable or a Keras variable.

Args
variable The variable to update.
value The value to add to the variable.

assign_sub

View source

assign_sub(
    variable, value
)

Subtract a value from a variable.

This should be used in optimizers instead of variable.assign_sub(value) to support backend specific optimizations. Note that the variable can be a model variable or an optimizer variable; it can be a backend native variable or a Keras variable.

Args
variable The variable to update.
value The value to add to the variable.

build

View source

build(
    var_list
)

Initialize optimizer variables.

Adam optimizer has 3 types of variables: momentums, velocities and velocity_hat (only set when amsgrad is applied),

Args
var_list list of model variables to build Adam variables on.

exclude_from_weight_decay

View source

exclude_from_weight_decay(
    var_list=None, var_names=None
)

Exclude variables from weight decay.

This method must be called before the optimizer's build method is called. You can set specific variables to exclude out, or set a list of strings as the anchor words, if any of which appear in a variable's name, then the variable is excluded.

Args
var_list A list of Variables to exclude from weight decay.
var_names A list of strings. If any string in var_names appear in the model variable's name, then this model variable is excluded from weight decay. For example, var_names=['bias'] excludes all bias variables from weight decay.

finalize_variable_values

View source

finalize_variable_values(
    var_list
)

Set the final value of model's trainable variables.

Sometimes there are some extra steps before ending the variable updates, such as overriding the model variables with its average value.

Args
var_list list of model variables.

from_config

View source

@classmethod
from_config(
    config, custom_objects=None
)

Creates an optimizer from its config.

This method is the reverse of get_config, capable of instantiating the same optimizer from the config dictionary.

Args
config A Python dictionary, typically the output of get_config.
custom_objects A Python dictionary mapping names to additional user-defined Python objects needed to recreate this optimizer.

Returns
An optimizer instance.

get_config

View source

get_config()

Returns the config of the optimizer.

An optimizer config is a Python dictionary (serializable) containing the configuration of an optimizer. The same optimizer can be reinstantiated later (without any saved state) from this configuration.

Subclass optimizer should override this method to include other hyperparameters.

Returns
Python dictionary.

load_own_variables

View source

load_own_variables(
    store
)

Set the state of this optimizer object.

lr_adjust

View source

lr_adjust(
    x
)

Adjusts learning rate based on the Moonlight implementation. This method enhances the stability of Muon, allowing it to use the same learning rate and weight decay as Adam. For details, see https://arxiv.org/abs/2502.16982 For a 2D matrix, the update is scaled by sqrt(max(n, m)) * rms_rate, where n and m are the dimensions of the matrix.

save_own_variables

View source

save_own_variables(
    store
)

Get the state of this optimizer object.

scale_loss

View source

scale_loss(
    loss
)

Scale the loss before computing gradients.

Scales the loss before gradients are computed in a train_step. This is primarily useful during mixed precision training to prevent numeric underflow.

set_weights

View source

set_weights(
    weights
)

Set the weights of the optimizer.

stateless_apply

View source

stateless_apply(
    optimizer_variables, grads, trainable_variables
)

Stateless version of apply that returns modified variables.

Args
optimizer_variables list of tensors containing the current values for the optimizer variables. These are native tensors and not keras.Variables.
grads list of gradients to apply.
trainable_variables list of tensors containing the current values for the model variables. These are native tensors and not keras.Variables.

Returns: A tuple containing two list of tensors, the updated trainable_variables and the updated optimizer_variables.

transpose_last_axis

View source

transpose_last_axis(
    X
)

update_step

View source

update_step(
    gradient, variable, learning_rate
)

zeropower_via_newtonschulz5

View source

zeropower_via_newtonschulz5(
    x, steps: int
)

We apply the Newton-Schulz iteration to compute matrix G.

We select a quintic iteration that maximizes the slope at zero. This approach helps minimize steps, even if the iteration doesn't fully converge across the interval. The result isn't exactly UV^T (from the SVD of G), but rather an approximation like US'V^T. Despite this approximation, model performance remains unaffected compared to using the exact UV^T from the SVD.