Struct RestoredWeightedTraining 

pub struct RestoredWeightedTraining<M, O, S, U> {
    pub model: M,
    pub optimizer: O,
    pub scheduler: S,
    pub accumulator: WeightedGradientsAccumulator<M>,
    pub state: U,
}

Available on crate feature std only.

Training components restored with actual unequal-microbatch normalization state.

Fields

model: M

Model with the recorded IDs and optional per-parameter storage dtypes.

optimizer: O

Recorded optimizer state, without a parameter update.

scheduler: S

Scheduler at the saved position, not advanced during restore.

accumulator: WeightedGradientsAccumulator<M>

Pending gradients, effective weight, microbatch count and loss scale.

state: U

Caller-owned source/sampler/RNG state from the same training checkpoint.