Struct LoraAttnWeights 

pub struct LoraAttnWeights {
    pub layer_idx: usize,
    pub q: Option<LoraPair>,
    pub k: Option<LoraPair>,
    pub v: Option<LoraPair>,
    pub o: Option<LoraPair>,
    pub kernels: LoraKernels,
    pub q_route: Option<LoraRoute>,
    pub k_route: Option<LoraRoute>,
    pub v_route: Option<LoraRoute>,
    pub o_route: Option<LoraRoute>,
}

Per-layer attention-side LoRA weights, installed by copy onto Qwen3AttentionLayer.

The q pair folds the delta into the RAW q_proj output at offset 0, full width = q_proj_dim (on a gated model the interleaved [Q|gate], width 2·q_heads·head_dim), BEFORE the deinterleave_qg split — the PEFT lora_B was trained against exactly that interleaved basis, so the delta applies like k/v/o, just wider.

Fields

layer_idx: usize

#30: the TRUE global layer index (0..num_hidden_layers), stamped at install from the global idx. The prefill apply sites index the request slot’s GLOBAL-layer-indexed pairs with THIS (not attn_layer_idx, an attention-only counter that diverges from the global index on hybrid GDN/attention models).

q: Option<LoraPair>

k: Option<LoraPair>

v: Option<LoraPair>

o: Option<LoraPair>

kernels: LoraKernels

q_route: Option<LoraRoute>

M2 per-request routing tables (per module). None = single/global adapter with no routing (the n==1 path uses the pair above and stays byte-identical). Some when a multi-adapter pool is resident; the batched decode path reads these + the per-seq seq_slot via the bgmv.

k_route: Option<LoraRoute>

v_route: Option<LoraRoute>

o_route: Option<LoraRoute>

Trait Implementations

impl Clone for LoraAttnWeights

fn clone(&self) -> LoraAttnWeights

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for LoraAttnWeights