Struct CapturedTensor 

pub struct CapturedTensor {
    pub name: String,
    pub layer: Option<usize>,
    pub ne0: usize,
    pub ne1: usize,
    pub data: Vec<f32>,
}

A single tensor copied out of the decode graph.

Produced by TensorCapture after a successful crate::LlamaContext::decode. For layer outputs ("l_out-N"), Self::layer is set to N.

Fields

name: String

Graph node name (e.g. "l_out-13", "result_norm").

layer: Option<usize>

Layer index when name is "l_out-{N}", otherwise None.

ne0: usize

First dimension (typically n_embd / hidden size).

ne1: usize

Second dimension (typically number of tokens in the batch position).

data: Vec<f32>

Flattened ne0 * ne1 values in ggml row-major order.

Index as data[token_idx * ne0 + dim_idx].

Implementations

impl CapturedTensor

pub fn n_embd(&self) -> usize

Number of embedding dimensions (alias for Self::ne0).

pub fn n_tokens(&self) -> usize

Number of token positions (alias for Self::ne1).

pub fn token_embedding(&self, token_idx: usize) -> Option<&[f32]>

Hidden-state vector for one token index.

Returns None when token_idx >= n_tokens().

Trait Implementations

impl Clone for CapturedTensor

fn clone(&self) -> CapturedTensor

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CapturedTensor

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

Formats the value using the given formatter.