Struct MoeExpertRoute 

pub struct MoeExpertRoute {
    pub a_table: DevicePtr,
    pub b_table: DevicePtr,
    pub scale_table: DevicePtr,
    pub n_experts: u32,
    pub k_in: u32,
    pub n_out: u32,
    pub max_rank: u32,
}

Per-EXPERT routing tables for the grouped down fold — the expert-keyed analogue of the slot-keyed super::lora_delta::LoraRoute. Built once at adapter install from the layer’s Down pairs; load-time-fixed device addresses, so they are stable kernel args across capture/replay (adapter identity for a mixed batch flows through the per-row moe_row_adapter, not these tables).

n_experts is the TABLE LENGTH = max adapted expert id + 1 (NOT the layer’s full num_experts): the grid launches grid.z = n_experts, and every adapted expert has index < n_experts, so any higher-index expert is unadapted and correctly folds nothing. expert_offsets[e] / [e+1] are read for e < n_experts <= num_experts, always in range of the [num_experts+1] prefix sum.

Fields

a_table: DevicePtr

[n_experts] u64 device array of A_e addresses (0 = expert unadapted).

b_table: DevicePtr

[n_experts] u64 device array of B_e addresses (0 = expert unadapted).

scale_table: DevicePtr

[n_experts] f32 device array of per-expert scale_e (0.0 where unadapted).

n_experts: u32

Table length = max adapted expert id + 1 (== grid.z).

k_in: u32

Contraction dim of the shrink stage (moe_intermediate_size).

n_out: u32

Output dim of the expand stage (hidden_size).

max_rank: u32

Padded rank (contraction dim of the expand stage; row stride of B_e).

Trait Implementations

impl Clone for MoeExpertRoute

fn clone(&self) -> MoeExpertRoute

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MoeExpertRoute

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

Formats the value using the given formatter.

impl Copy for MoeExpertRoute