Struct GradientTape 

pub struct GradientTape { /* private fields */ }

Gradient tape for reverse-mode automatic differentiation.

Usage

1. Create a tape.
2. Record operations during the forward pass using record_* methods. Each method returns an output tensor index.
3. After computing the scalar loss, call backward with the gradient of the loss w.r.t. the output to accumulate gradients into all upstream tensors.

Implementations

impl GradientTape

pub fn new() -> Self

Create an empty tape.

pub fn record_add(&mut self, a: usize, b: usize) -> usize

Record an elementwise add: out = tensors[a] + tensors[b].

Returns the output tensor index.

pub fn record_mul( &mut self, a: usize, a_data: &Array1<f32>, b: usize, b_data: &Array1<f32>, ) -> usize

Record an elementwise multiply: out = tensors[a] * tensors[b].

a_data and b_data are the forward values saved for gradient computation (product rule).

Returns the output tensor index.

pub fn record_matmul( &mut self, a: usize, a_mat: &Array2<f32>, b: usize, b_mat: &Array2<f32>, ) -> usize

Record a matrix-matrix multiply: out = A @ B (result flattened to 1-D).

Returns the output tensor index.

pub fn record_silu(&mut self, input: usize, input_data: &Array1<f32>) -> usize

Record a SiLU activation: out = x * sigmoid(x).

input_data is the pre-activation tensor value.

Returns the output tensor index.

pub fn record_layer_norm( &mut self, input: usize, mean: f32, var: f32, scale: &Array1<f32>, ) -> usize

Record a layer-norm operation.

Returns the output tensor index.

pub fn record_ssm_scan( &mut self, input: usize, a_vals: &Array1<f32>, b_vals: &Array1<f32>, ) -> usize

Record a simplified SSM scan step.

Returns the output tensor index.

pub fn backward( &self, loss_grad: Array1<f32>, tensors: &mut Vec<Array1<f32>>, ) -> ModelResult<()>

Run reverse-mode backpropagation.

Parameters

• loss_grad: gradient of the scalar loss w.r.t. the final output (shape must match the last recorded output).
• tensors: gradient buffers, one Array1<f32> per allocated tensor slot. The caller is responsible for initialising these to zeros and ensuring tensors.len() >= self.num_tensors. After the call each buffer holds the accumulated gradient for that tensor slot.

Errors

Returns ModelError::numerical_instability if any gradient contains NaN or Inf values.

Trait Implementations

impl Default for GradientTape

fn default() -> Self

Returns the “default value” for a type.