Struct GraphAttnGeom 

pub struct GraphAttnGeom<'a> {
    pub nkv: usize,
    pub hd: usize,
    pub dv: usize,
    pub rd: usize,
    pub invf: &'a [f32],
    pub rope_scale: f32,
    pub window: Option<usize>,
    pub sink: Option<&'a [f32]>,
    pub v_norm: bool,
}

Per-layer attention geometry for the wgpu graphs (see GraphAttn::Full::geom). The layer’s CPU cache keeps K rows hd wide and V rows zero-padded to hd; the device mirror stores V dv wide, and a windowed layer keeps a ring of the last positions only.

Fields

nkv: usize

KV heads of this layer (divides the Q heads).

hd: usize

Q/K head width of this layer. The call-wide hd for every model but Gemma-4, whose global layers are 512 wide over 256-wide sliding ones (<= 512, a multiple of 4).

dv: usize

V head width, 4 <= dv <= hd, a multiple of 4.

rd: usize

Rotary width of this layer (NeoX half-split over [0, rd)).

invf: &'a [f32]

This layer’s RoPE inverse frequencies (rd / 2 of them).

rope_scale: f32

YaRN attention-amplitude factor applied after the rotary pair transform. 1.0 is ordinary RoPE; Mellum’s global layers use 1.277258872…. This is deliberately per-layer: local and global layers may have different RoPE tables and factors.

window: Option<usize>

Positions a query sees, its own included (MiMo-V2 SWA: 128); None = the whole context.

sink: Option<&'a [f32]>

Learned per-Q-head sink logits (gpt-oss / MiMo-V2): they join the softmax max and denominator and carry no value row.

v_norm: bool

Gemma-4’s scale-less V norm: every V head is RMS-normalized (no weight) before it is cached, as attention::vnorm_head does.

Trait Implementations

impl<'a> Clone for GraphAttnGeom<'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 GraphAttnGeom<'a>