Struct ZBlockRef 

pub struct ZBlockRef<'a> {

    pub wq: usize,
    pub wk: usize,
    pub wv: usize,
    pub wo: usize,
    pub w1: usize,
    pub w3: usize,
    pub w2: usize,
    pub norm1: &'a [f32],
    pub norm2: &'a [f32],
    pub ffn_norm1: &'a [f32],
    pub ffn_norm2: &'a [f32],
    pub norm_q: &'a [f32],
    pub norm_k: &'a [f32],
}

One Z-Image transformer block’s device inputs (noise refiner, context refiner or main layer — all share this shape). Weights are tensor indices into model.tensors (diffusers names under dit.); the codec is whatever the container holds (F16/Bf16/Q8Row/Q8_2f/Q4TiledP…), and a backend that cannot expand a codec declines (returns false). Norm vectors are f32 host slices that live as long as the caller’s ZImageDit; a backend may cache them by pointer (they do not change).

Fields

wq: usize

attention.to_q/to_k/to_v/to_out.0.weight, each [hidden, hidden].

wk: usize

wv: usize

wo: usize

w1: usize

feed_forward.w1 (gate) / w3 (up) [inter, hidden], w2 (down) [hidden, inter]. FFN = w2(silu(w1·x) ⊙ w3·x).

w3: usize

w2: usize

norm1: &'a [f32]

attention_norm1 / attention_norm2, [hidden] (plain-w RMSNorm).

norm2: &'a [f32]

ffn_norm1: &'a [f32]

ffn_norm1 / ffn_norm2, [hidden].

ffn_norm2: &'a [f32]

norm_q: &'a [f32]

attention.norm_q / norm_k, [hd] (per-head RMSNorm before RoPE).

norm_k: &'a [f32]

Trait Implementations

impl<'a> Clone for ZBlockRef<'a>

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl<'a> Copy for ZBlockRef<'a>