Struct LoraLinear 

pub struct LoraLinear { /* private fields */ }

A single LoRA-adapted linear layer.

Computes: output = W @ x + (alpha/rank) * B @ (A @ x)

• lora_a has shape (rank, in_features) — initialized with Kaiming uniform
• lora_b has shape (out_features, rank) — initialized with zeros

Because B starts at zero, the initial LoRA contribution is zero and the model produces the same output as the original frozen weights.

Implementations

impl LoraLinear

pub fn new(weight: Array2<f32>, rank: usize, alpha: f32) -> ModelResult<Self>

Create a new LoRA-adapted linear layer.

The weight matrix should have shape (out_features, in_features). LoRA A is initialized with Kaiming-uniform values, B with zeros.

pub fn forward(&self, input: &Array1<f32>) -> ModelResult<Array1<f32>>

Forward pass for a single input vector.

Computes: output = W @ x + scaling * B @ (A @ x)

pub fn forward_batch(&self, input: &Array2<f32>) -> ModelResult<Array2<f32>>

Forward pass for a batch of inputs.

Input shape: (batch_size, in_features) Output shape: (batch_size, out_features)

pub fn merge(&mut self) -> ModelResult<()>

Merge LoRA weights into the original weight matrix for inference.

After merging, forward passes use only the modified W with no extra computation. W = W + scaling * B @ A

pub fn unmerge(&mut self) -> ModelResult<()>

Unmerge LoRA weights from the original weight matrix.

Restores W to its original values for continued training. W = W - scaling * B @ A

pub fn trainable_params(&self) -> usize

Number of trainable parameters: rank * (in_features + out_features)

pub fn total_params(&self) -> usize

Total parameters including frozen weights

pub fn compression_ratio(&self) -> f32

Ratio of trainable to total parameters

pub fn lora_a(&self) -> &Array2<f32>

Get reference to LoRA A matrix

pub fn lora_b(&self) -> &Array2<f32>

Get reference to LoRA B matrix

pub fn set_lora_a(&mut self, a: Array2<f32>) -> ModelResult<()>

Set the LoRA A matrix, validating dimensions

pub fn set_lora_b(&mut self, b: Array2<f32>) -> ModelResult<()>

Set the LoRA B matrix, validating dimensions

pub fn enable(&mut self)

Enable LoRA adaptation

pub fn disable(&mut self)

Disable LoRA adaptation (output equals original W @ x)

pub fn is_enabled(&self) -> bool

Whether LoRA is currently enabled

pub fn is_merged(&self) -> bool

Whether LoRA weights are merged into W

pub fn weight(&self) -> &Array2<f32>

Get the weight matrix reference

pub fn rank(&self) -> usize

Get the rank

pub fn alpha(&self) -> f32

Get the alpha

pub fn scaling(&self) -> f32

Get the scaling factor

Trait Implementations

impl Clone for LoraLinear

fn clone(&self) -> LoraLinear

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LoraLinear

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

Formats the value using the given formatter.