Struct MoeJob 

pub struct MoeJob<'a> {
    pub gate: (usize, usize, usize, &'a [f32]),
    pub up: (usize, usize, usize, &'a [f32]),
    pub down: (usize, usize, usize, &'a [f32]),
    pub xs_gate: Vec<f32>,
    pub xs_up: Vec<f32>,
    pub down_col: &'a [f32],
    pub w: f32,
    pub q1: bool,
    pub q4t: bool,
    pub q4tp: bool,
    pub gu_q2: bool,
    pub swiglu_limit: f32,
}

A single MoE-FFN job (an expert with its own weight), executed in one submission: (rows, cols, idx, row_scale) for gate/up/down + prescaled inputs + the down θ-field + the blending weight.

Fields

gate: (usize, usize, usize, &'a [f32])

up: (usize, usize, usize, &'a [f32])

down: (usize, usize, usize, &'a [f32])

xs_gate: Vec<f32>

xs_up: Vec<f32>

down_col: &'a [f32]

w: f32

q1: bool

q1 trio: scales live inside the 6-byte tiles (row_scale slices empty, xs raw f32). Backends without a q1 kernel refuse the job.

q4t: bool

q4_tiled trio: scales inside the 18-byte tiles (row_scale slices empty, xs raw f32) — the MoE-hybrid coder class.

q4tp: bool

q4tp trio: same raw-xs contract, 16-byte nibble stride and the scale on a per-row ladder. Without this the experts of a q4tp MoE model fall to the CPU while every other dtype rides the device.

gu_q2: bool

Mixed 2-bit profile: gate/up are q2tp (8-byte chunks, zero rung), down stays q4tp. Set together with q4tp; a backend without the 2-bit kernel must refuse the whole job.

swiglu_limit: f32

The reference’s swiglu_limit; 0 disables the clamp. A backend that cannot apply it must REFUSE the job rather than drop it silently — the difference only shows on saturating activations, which is the hardest kind of divergence to notice.