Struct LoraStack 

pub struct LoraStack { /* private fields */ }

A stack of LoRA adapters applied in order.

Adapters are applied additively: output += scale_0 · LoRA_0(x) + scale_1 · LoRA_1(x) + ...

Each adapter’s intrinsic alpha/rank scale is multiplied by the per-entry stack_scale.

The stack supports two independent adapter lists:

• entries: legacy LoadedLora adapters (GGUF-loaded, keyed by tensor name).
• adapter_list: trait-object adapters implementing LoraAdapterTrait.

Both lists are consulted by the respective dispatch methods.

Implementations

impl LoraStack

pub fn new() -> LoraStack

Create an empty stack.

pub fn push(&mut self, adapter: Arc<LoadedLora>, scale: f32)

Push a new GGUF-loaded adapter onto the stack with a given scale multiplier.

scale multiplies the adapter’s intrinsic alpha/rank scale factor.

pub fn pop(&mut self) -> Option<(Arc<LoadedLora>, f32)>

Remove the last GGUF-loaded adapter from the stack.

Returns None if the entries list is empty.

pub fn clear(&mut self)

Remove all adapters (both lists).

pub fn len(&self) -> usize

Number of stacked legacy adapters.

pub fn is_empty(&self) -> bool

Returns true if no adapters are stacked in either list.

pub fn entries(&self) -> &[(Arc<LoadedLora>, f32)]

Immutable view of the (LoadedLora, scale) entries.

pub fn adapters(&self) -> &[Arc<dyn LoraAdapterTrait>]

Immutable view of the trait-object adapter list.

pub fn apply( &self, tensor_name: &str, input: &[f32], out_features: usize, ) -> Result<Vec<f32>, ArchError>

Apply the entire legacy stack to compute the additive LoRA delta for tensor_name.

Returns Ok(delta) where delta is a Vec<f32> of length out_features representing Σ_i stack_scale_i · (alpha_i/rank_i) · B_i @ A_i @ input. Adapters that do not define tensor_name are silently skipped. The caller is responsible for adding delta to the base linear output.

Errors

Returns crate::error::ArchError::Quant if a matching adapter’s dimension checks fail.

pub fn push_adapter(&mut self, adapter: Arc<dyn LoraAdapterTrait>)

Push a trait-object adapter onto the new adapter list.

pub fn applied_delta( &self, target: TargetModule, layer: usize, input: &[f32], ) -> Option<Vec<f32>>

Compute the combined delta for (target, layer) across all trait-object adapters, applied to input vector input.

Returns None if no adapter in the list covers this (target, layer).

Formula: Δ_total = Σᵢ (αᵢ / rᵢ) * Bᵢ @ Aᵢ @ input.

Trait Implementations

impl Clone for LoraStack

fn clone(&self) -> LoraStack

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LoraStack

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

Formats the value using the given formatter.

impl Default for LoraStack

fn default() -> LoraStack

Returns the “default value” for a type.