Struct AccumulationConfig 

pub struct AccumulationConfig {
    pub accumulation_steps: usize,
    pub normalize: bool,
    pub max_grad_norm: Option<f64>,
}

Configuration for gradient accumulation.

Fields

accumulation_steps: usize

Number of micro-batches to accumulate before triggering an update.

normalize: bool

Whether to normalize (average) gradients across micro-batches.

max_grad_norm: Option<f64>

Maximum gradient norm for clipping (None = no clipping).

Implementations

impl AccumulationConfig

pub fn new(steps: usize) -> Self

Create a new config with the given number of accumulation steps. Clamps to a minimum of 1.

pub fn with_normalize(self, normalize: bool) -> Self

Set whether to normalize (average) gradients.

pub fn with_max_grad_norm(self, norm: f64) -> Self

Set maximum gradient norm for clipping.

pub fn effective_batch_size(&self, micro_batch_size: usize) -> usize

Compute the effective batch size given the micro-batch size.

Trait Implementations

impl Clone for AccumulationConfig

fn clone(&self) -> AccumulationConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AccumulationConfig

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

Formats the value using the given formatter.

impl Default for AccumulationConfig

fn default() -> Self

Returns the “default value” for a type.