Struct LoraLayerWeights 

pub struct LoraLayerWeights {
    pub layer_idx: usize,
    pub q_proj: Option<LoraPair>,
    pub k_proj: Option<LoraPair>,
    pub v_proj: Option<LoraPair>,
    pub o_proj: Option<LoraPair>,
    pub gate_proj: Option<LoraPair>,
    pub up_proj: Option<LoraPair>,
    pub down_proj: Option<LoraPair>,
    pub out_proj: Option<LoraPair>,
    pub router: Option<LoraPair>,
    pub experts: Option<ExpertLoraLayer>,
}

One full-attention layer’s adapted modules. None = module not adapted. Pairs are the CANONICAL LoraPair from layers::ops::lora_delta (Copy — installed by copy into the layer structs at model build).

Clone (LoraPair is Copy) so a slot’s layers can be re-installed onto the layer structs on a runtime rotation (set_active_lora).

Fields

layer_idx: usize

q_proj: Option<LoraPair>

k_proj: Option<LoraPair>

v_proj: Option<LoraPair>

o_proj: Option<LoraPair>

gate_proj: Option<LoraPair>

up_proj: Option<LoraPair>

down_proj: Option<LoraPair>

out_proj: Option<LoraPair>

GDN out_proj delta (linear-attention layers only).

router: Option<LoraPair>

Feature-1: MoE router (mlp.gate) delta on the routing logits. None unless the adapter targets the router AND ATLAS_LORA_EXPERTS=1.

experts: Option<ExpertLoraLayer>

Feature-1: this layer’s routed-expert LoRA coverage (sparse per-expert pairs). None for attention/dense-only adapters or when expert LoRA is disabled. Packed into a SEPARATE expert allocation (not the equal-slot attention pool), so the BGMV route tables + slot offset math are untouched.

Implementations

impl LoraLayerWeights

pub fn empty(layer_idx: usize) -> Self

A zeroed per-layer entry (all modules unadapted). Used by the pack loops to lazily create a layer entry the first time any module — attention, router, or expert — lands on it.

impl LoraLayerWeights

pub fn module_pair(&self, module: LoraModule) -> Option<&LoraPair>

Select this layer’s LoraPair for module (None = module not adapted). The single source of truth for the (layer, module) → pair map, shared by select_routed_pair and LoraWeights::refresh_slot_tables.

Trait Implementations

impl Clone for LoraLayerWeights

fn clone(&self) -> LoraLayerWeights

Returns a duplicate of the value.

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

Performs copy-assignment from source.