Struct Placement 

pub struct Placement<'f> {
    pub row: &'f dyn Fn(usize, usize) -> usize,
    pub rope_theta: Option<f32>,
    pub decode_only: bool,
}

Where a stream’s output goes: element index of (head, token) row start and the token stride, plus optional RoPE (K given pre-RoPE).

Fields

row: &'f dyn Fn(usize, usize) -> usize

(head, token) -> element offset of that row’s channel 0.

rope_theta: Option<f32>

Apply half-split RoPE with this theta to each decoded token row.

decode_only: bool

Decode without writing the output (benchmark: the cost a fused decode + attention kernel cannot avoid).