Struct FullActivationCache 

pub struct FullActivationCache { /* private fields */ }

Stores all-position activations from a forward pass.

Unlike ActivationCache which stores only the last-token activation per layer, this cache stores the full residual stream at every token position. Each tensor has shape [seq_len, d_model].

Example

use candle_mi::FullActivationCache;
use candle_core::{Device, Tensor};

let mut cache = FullActivationCache::with_capacity(32);
// shape [seq_len=10, d_model=128]
cache.push(Tensor::zeros((10, 128), candle_core::DType::F32, &Device::Cpu).unwrap());

// Get a single position's activation for CLT encoding
let act = cache.get_position(0, 5).unwrap(); // shape [d_model]

Implementations

impl FullActivationCache

pub fn with_capacity(n_layers: usize) -> Self

Create an empty cache with capacity for n_layers layers.

pub fn push(&mut self, tensor: Tensor)

Add a layer’s all-position activation to the cache.

The tensor should have shape [seq_len, d_model].

pub fn get_layer(&self, layer: usize) -> Option<&Tensor>

Get the full activation tensor for a specific layer.

Returns shape [seq_len, d_model], or None if the layer is not in the cache.

pub fn get_position(&self, layer: usize, position: usize) -> Result<Tensor>

Get the activation at a specific layer and token position.

Returns shape [d_model] — compatible with CLT encode().

Errors

Returns MIError::Hook if the layer is not in the cache or the position is out of range.

pub const fn n_layers(&self) -> usize

Number of cached layers.

pub fn seq_len(&self) -> Result<usize>

Sequence length (from the first layer’s tensor).

Errors

Returns MIError::Hook if the cache is empty.

pub const fn is_empty(&self) -> bool

Whether the cache is empty.

Trait Implementations

impl Debug for FullActivationCache

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

Formats the value using the given formatter.