Struct AttnStep 

pub struct AttnStep {
    pub layer: usize,
    pub n_head: usize,
    pub q_len: usize,
    pub kv_len: usize,
    pub probs: Vec<f32>,
}

One block’s attention probabilities for one decode step, read back from the device.

probs is the row-major [n_head, q_len, kv_len] tensor the model actually attended with — captured after the softmax, not recomputed — so a visualization built from it shows the same arithmetic that produced the text. ops::softmax is out-of-place, so capturing it perturbs nothing.

Fields

layer: usize

n_head: usize

q_len: usize

Query positions in this step: the whole prompt on prefill, 1 per decode.

kv_len: usize

Past positions attended to, including the queries themselves.

probs: Vec<f32>

Trait Implementations

impl Clone for AttnStep

fn clone(&self) -> AttnStep

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AttnStep

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.