Struct GradientsAccumulator 

pub struct GradientsAccumulator<M> { /* private fields */ }

Accumulate gradients into a single Gradients object.

Implementations

impl<M> GradientsAccumulator<M>

pub fn new() -> Self

Create a new gradients accumulator.

impl<M> GradientsAccumulator<M>

pub fn try_to_record<B: AutodiffBackend>( &self, ) -> Result<GradientsParamsRecord, RecorderError>

where M: AutodiffModule<B>,

Snapshot pending gradients without resetting the accumulation window.

pub async fn to_record_async<B: AutodiffBackend>( &self, ) -> Result<GradientsParamsRecord, RecorderError>

where M: AutodiffModule<B>,

Asynchronously snapshot pending gradients without resetting the accumulator.

pub fn load_record<B: AutodiffBackend>( &mut self, record: GradientsParamsRecord, device: &B::Device, ) -> Result<(), RecorderError>

where M: AutodiffModule<B>,

Replace pending gradients with checkpoint state on the given device.

The model IDs and the caller’s accumulation count must be restored from the same checkpoint. A rejected record leaves the accumulator unchanged.

pub fn accumulate<B: AutodiffBackend>( &mut self, module: &M, grads: GradientsParams, )

where M: AutodiffModule<B>,

Accumulate the given gradients for each parameter in the given module.

pub fn accumulate_with_dtype<B: AutodiffBackend>( &mut self, module: &M, grads: GradientsParams, dtype: FloatDType, )

where M: AutodiffModule<B>,

Accumulate after converting incoming and pending gradients to the given dtype.

FP32 accumulation can retain small additions to half-precision gradients. Parameter storage, loss normalization and accumulation counts are unchanged. Checkpoints preserve the pending gradient dtype; select the same dtype on subsequent calls after restoring. No accumulator reset is performed.

pub fn grads(&mut self) -> GradientsParams

Return the accumulated gradients and reset the accumulator state.

Trait Implementations

impl<M> Default for GradientsAccumulator<M>

fn default() -> Self

Returns the “default value” for a type.