Struct ModelStateTrainingRecord 

pub struct ModelStateTrainingRecord<B, M, O, S, R, U>
where
    B: AutodiffBackend,
    M: AutodiffModule<B>,
    O: Optimizer<M, B>,
    S: LrScheduler,
    R: Record<B>,
    U: Record<B>,
{ /* private fields */ }

Available on crate feature std only.

Caller-selected model state plus actual optimizer/scheduler/pending gradients.

R can be RUDA’s A/B-only native adapter record. It must restore every trainable parameter, including IDs/dtypes, and identify any omitted frozen state. Capture R and these components at the same boundary, without concurrent updates. No full model record or hidden frozen tensor copy is created here.

Implementations

impl<B, M, O, S, R, U> ModelStateTrainingRecord<B, M, O, S, R, U>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>, S: LrScheduler, R: Record<B>, U: Record<B>,

pub fn capture( model: &M, model_state: R, optimizer: &O, scheduler: &S, accumulator: &GradientsAccumulator<M>, state: U, ) -> Result<Self, RecorderError>

Capture provided actual model state, optimizer, scheduler and pending values.

pub async fn capture_async( model: &M, model_state: R, optimizer: &O, scheduler: &S, accumulator: &GradientsAccumulator<M>, state: U, ) -> Result<Self, RecorderError>

Same capture with asynchronous readback of actual pending gradients.

pub fn capture_weighted( model: &M, model_state: R, optimizer: &O, scheduler: &S, accumulator: &WeightedGradientsAccumulator<M>, state: U, ) -> Result<ModelStateTrainingRecord<B, M, O, S, R, (WeightedAccumulationState, U)>, RecorderError>

Capture actual weighted-window counts/scale alongside caller source/RNG state.

pub async fn capture_weighted_async( model: &M, model_state: R, optimizer: &O, scheduler: &S, accumulator: &WeightedGradientsAccumulator<M>, state: U, ) -> Result<ModelStateTrainingRecord<B, M, O, S, R, (WeightedAccumulationState, U)>, RecorderError>

Asynchronous weighted capture, with no implicit window reset or update.

pub fn save<C: Recorder<B>>( self, recorder: &C, args: C::RecordArgs, ) -> Result<C::RecordOutput, RecorderError>

Save provided model state and live training components in one recorder payload.

pub fn load<C: Recorder<B>>( recorder: &C, args: C::LoadArgs, device: &B::Device, ) -> Result<Self, RecorderError>

Read the combined state on the requested device; recreate omitted base state separately.

pub fn restore<F>( self, model: M, optimizer: O, scheduler: S, device: &B::Device, restore_model: F, ) -> Result<RestoredTraining<M, O, S, U>, RecorderError>

where F: FnOnce(R, M) -> Result<M, RecorderError>,

Restore caller-selected model state first, then validate actual trainable identities. restore_model may call a native adapter record’s restore_into with an independently supplied frozen base identity. The prepared model must already use the requested device; the omitted base is never copied or moved implicitly.

impl<B, M, O, S, R, U> ModelStateTrainingRecord<B, M, O, S, R, (WeightedAccumulationState, U)>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>, S: LrScheduler, R: Record<B>, U: Record<B>,

pub fn restore_weighted<F>( self, model: M, optimizer: O, scheduler: S, device: &B::Device, restore_model: F, ) -> Result<RestoredWeightedTraining<M, O, S, U>, RecorderError>

where F: FnOnce(R, M) -> Result<M, RecorderError>,

Restore actual A/B/model state and the weighted accumulation window together.

Trait Implementations

impl<B, M, O, S, R, U> Record<B> for ModelStateTrainingRecord<B, M, O, S, R, U>

where B: AutodiffBackend, M: AutodiffModule<B>, O: Optimizer<M, B>, S: LrScheduler, R: Record<B>, U: Record<B>,

type Item<P: PrecisionSettings> = (u32, <R as Record<B>>::Item<P>, <TrainableParameterContract as Record<B>>::Item<P>, <<O as Optimizer<M, B>>::Record as Record<B>>::Item<P>, <<S as LrScheduler>::Record<B> as Record<B>>::Item<P>, <GradientsParamsRecord as Record<B>>::Item<P>, <U as Record<B>>::Item<P>)

Type of the item that can be serialized and deserialized.

fn into_item<P: PrecisionSettings>(self) -> Self::Item<P>

Convert the current record into the corresponding item that follows the given settings.

fn from_item<P: PrecisionSettings>( item: Self::Item<P>, device: &B::Device, ) -> Self

Convert the given item into a record.