Struct DevExps 

pub struct DevExps {
    pub gate: CudaSlice<u8>,
    pub up: CudaSlice<u8>,
    pub down: CudaSlice<u8>,
    pub ptr_row: CudaSlice<u64>,
    pub gu_il: bool,
}

Device-resident expert slabs for one layer (gate/up/down) + the prebuilt [3, n_expert] pointer row the _dev kernels consume.

Fields

gate: CudaSlice<u8>

up: CudaSlice<u8>

down: CudaSlice<u8>

ptr_row: CudaSlice<u64>

[3*n_expert] u64 device row: gate ptrs, up ptrs, down ptrs (proj-major like layer_dev_row).

gu_il: bool

WALL-GAP ARC (MEMRA_MOE_GU_IL=1): gate/up rows INTERLEAVED in one slab — row o of gate at base + o*(rb_g+rb_u), up at +rb_g. Consumers on the dev path must use (rb_g+rb_u) as the row stride for BOTH projections (see MoeWeights::dev_rb_gu). One contiguous 1760B stream per (expert,row) instead of two scattered 880B streams — the measured 56%-of-wall fix candidate. Kernels unchanged (stride is already a parameter everywhere).