Struct LayerDetail 

pub struct LayerDetail {
    pub layer: usize,
    pub ln1_out: Vec<f32>,
    pub q: Vec<f32>,
    pub k: Vec<f32>,
    pub v: Vec<f32>,
    pub scores: Vec<f32>,
    pub attn_head_out: Vec<f32>,
    pub attn_proj_out: Vec<f32>,
    pub resid_attn: Vec<f32>,
    pub ln2_out: Vec<f32>,
    pub mlp_hidden: Vec<f32>,
    pub block_out: Vec<f32>,
}

One block’s Q/K/V, MLP and output tensors for one step, read back from the device. Captured only for the first detail_layers blocks — the expensive kinds — while AttnStep is captured for every block.

q/k/v are the tensors for the new positions only ([n_head, q_len, head_dim]), which is what the step computed: the older keys and values live in the cache and were never recomputed.

Fields

layer: usize

ln1_out: Vec<f32>

[q_len, n_embd], the block’s first LayerNorm.

q: Vec<f32>

[n_head, q_len, head_dim]

k: Vec<f32>

v: Vec<f32>

scores: Vec<f32>

[n_head, q_len, kv_len] — q · kᵢ / √head_dim before the causal mask and the softmax. The masked triangle only means something next to the full rectangle it was cut from.

attn_head_out: Vec<f32>

[q_len, n_embd] — the heads concatenated, before the output projection. Kept separate from attn_proj_out because this is the one place n_head × head_dim becomes n_embd.

attn_proj_out: Vec<f32>

[q_len, n_embd], after the attention output projection.

resid_attn: Vec<f32>

[q_len, n_embd], after the first residual add.

ln2_out: Vec<f32>

[q_len, n_embd], the block’s second LayerNorm.

mlp_hidden: Vec<f32>

[q_len, 4 * n_embd], post-GELU.

block_out: Vec<f32>

[q_len, n_embd], after the second residual add.

Trait Implementations

impl Clone for LayerDetail

fn clone(&self) -> LayerDetail

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LayerDetail

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

Formats the value using the given formatter.