Struct VitArgs 

pub struct VitArgs<'a> {

    pub key: u64,
    pub layers: Vec<VitLayer<'a>>,
    pub n: usize,
    pub d: usize,
    pub heads: usize,
    pub head_dim: usize,
    pub inter: usize,
    pub eps: f64,
    pub gelu_tanh: bool,
    pub rope_halves: bool,
    pub cos: &'a [f32],
    pub sin: &'a [f32],
    pub prof: bool,
}

A vision tower’s encoder stack for vit_encode. cos/sin are [n, head_dim] per lane; rope_halves pairs lane j with j ± hd/2 (Qwen-VL), else with its neighbour (OneVision’s interleaved form). key identifies the tower for the device weight cache.

Fields

key: u64

layers: Vec<VitLayer<'a>>

n: usize

d: usize

heads: usize

head_dim: usize

inter: usize

eps: f64

gelu_tanh: bool

rope_halves: bool

cos: &'a [f32]

sin: &'a [f32]

prof: bool