Struct LoRAAdapter 

pub struct LoRAAdapter {
    pub a: Tensor,
    pub b: Tensor,
    pub alpha: f32,
    pub rank: usize,
}

Low-rank adapter added to a Linear: x @ a @ b * alpha / rank.

Fields

a: Tensor

Down projection [in, rank].

b: Tensor

Up projection [rank, out] (zero-initialised, so training starts from the base model).

alpha: f32

Scale numerator; the update is scaled by alpha / rank.

rank: usize

Adapter rank.

Implementations

impl LoRAAdapter

pub fn new( ctx: Arc<WgpuContext>, in_dim: usize, out_dim: usize, rank: usize, alpha: f32, ) -> Self

An adapter for an in_dim -> out_dim projection. a is drawn uniformly from ±1/sqrt(in_dim) (Kaiming-uniform, as in PEFT) with a deterministic generator seeded from the shape, and b is zero.

A constant a would make every rank component identical, so the adapter could never learn more than a rank-1 update.

pub fn new_seeded( ctx: Arc<WgpuContext>, in_dim: usize, out_dim: usize, rank: usize, alpha: f32, seed: u64, ) -> Self

new with an extra seed, so adapters of equal shape (e.g. every layer’s q_proj) start from different matrices.

pub fn params(&self) -> Vec<Tensor>

The adapter tensors [a, b].

Trait Implementations

impl Clone for LoRAAdapter

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LoRAAdapter

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

Formats the value using the given formatter.