Struct TensorCapture 

pub struct TensorCapture { /* private fields */ }

Captures intermediate tensor outputs during llama_decode.

Create a TensorCapture, attach it to LlamaContextParams via with_tensor_capture, then call decode(). After decode completes, read captured data via [get], [get_layer], or [iter].

Lifetime & Safety

The TensorCapture must outlive the LlamaContext it is attached to. The borrow is enforced by with_tensor_capture taking &mut self.

Implementations

impl TensorCapture

pub fn for_layers(layer_indices: &[usize]) -> Self

Create a capture that intercepts layer outputs "l_out-{N}" for the specified layer indices.

This is the most common use case for extracting per-layer hidden states from a language model.

Example

// Capture layers 13, 20, 27 (typical for LLaMA-3.2-3B with positions [0.5, 0.75, 1.0])
let mut capture = TensorCapture::for_layers(&[13, 20, 27]);

pub fn for_names(names: &[&str]) -> Self

Create a capture that intercepts tensors with exact matching names.

Example

let mut capture = TensorCapture::for_names(&["result_norm", "l_out-27"]);

pub fn for_prefix(prefix: &str) -> Self

Create a capture that intercepts all tensors whose name starts with the given prefix.

Example

// Capture all attention outputs
let mut capture = TensorCapture::for_prefix("attn_out");

pub fn all() -> Self

Create a capture that intercepts all tensors.

⚠️ Warning: this can produce very large amounts of data. Use only for debugging or inspection.

pub fn clear(&mut self)

Clear all previously captured data, keeping the filter configuration.

Call this before a new decode() if reusing the capture across multiple batches.

pub fn get(&self, name: &str) -> Option<&CapturedTensor>

Get a captured tensor by its full name (e.g. "l_out-13").

pub fn get_layer(&self, layer_idx: usize) -> Option<&CapturedTensor>

Get a captured layer output by layer index.

Looks up "l_out-{layer_idx}".

pub fn has_layer(&self, layer_idx: usize) -> bool

Returns true if the specified layer was captured.

pub fn len(&self) -> usize

Number of tensors captured so far.

pub fn is_empty(&self) -> bool

Returns true if no tensors have been captured.

pub fn iter(&self) -> impl Iterator<Item = (&str, &CapturedTensor)>

Iterate over all captured tensors.

pub fn captured_layers(&self) -> Vec<usize>

Get all captured layer indices (sorted).

Trait Implementations

impl Debug for TensorCapture

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

Formats the value using the given formatter.