Skip to main content

DeltaReducer

Trait DeltaReducer 

Source
pub trait DeltaReducer {
    // Required method
    fn reduce_key(
        &self,
        key: &str,
        contributions: &[Contribution],
    ) -> Result<(Tensor, Tensor), Error>;
}
Expand description

Order-independent N-way delta reducer.

Implementors merge an explicit contributor set into a fixed-rank delta. The stack + truncated-SVD implementation is StackSvdReducer; the trait keeps it swappable (held-out-eval variants, future native reducers).

Required Methods§

Source

fn reduce_key( &self, key: &str, contributions: &[Contribution], ) -> Result<(Tensor, Tensor), Error>

Reduce contributors for a single layer key into fixed-rank factors (A_out [R, in], B_out [out, R]), with the result expressed at scaling = 1.0 (weights + per-contributor scaling already folded in).

All tensors stay on contributions[*].lora_a.device() (must be uniform).

Implementors§