Struct SparseGPTImportance 

pub struct SparseGPTImportance { /* private fields */ }

SparseGPT importance estimator using Hessian-based saliency.

Computes importance scores based on the Optimal Brain Surgeon (OBS) saliency metric: saliency = w^2 / H^{-1}_{jj}

This identifies weights whose removal causes minimal output perturbation when compensated by adjusting remaining weights.

Algorithm

1. Compute Hessian H = (1/n) * X^T * X + damp * I
2. Compute Hessian inverse via Cholesky decomposition
3. Saliency = w^2 / diag(H^{-1})

Key Insight

Second-order information allows weight updates that minimize the output perturbation caused by pruning.

Implementations

impl SparseGPTImportance

pub fn new(layer_name: impl Into<String>) -> SparseGPTImportance

Create SparseGPT importance estimator for a specific layer.

Arguments

• layer_name - Layer identifier to look up in CalibrationContext

pub fn with_block_size(self, block_size: usize) -> SparseGPTImportance

Set block size for block-wise processing.

Smaller blocks use less memory but may be less accurate.

pub fn with_damp(self, damp: f32) -> SparseGPTImportance

Set absolute damping factor.

Arguments

• damp - Value added to diagonal of Hessian for stability

pub fn with_relative_damp(self, damp: f32) -> SparseGPTImportance

Set relative damping factor.

Arguments

• damp - Percentage of mean diagonal to add (e.g., 0.01 = 1%)

pub fn layer_name(&self) -> &str

Get the layer name.

pub fn block_size(&self) -> usize

Get the block size.

pub fn damp(&self) -> f32

Get the damping factor.

pub fn compute_hessian( &self, activations: &Tensor, ) -> Result<Tensor, PruningError>

Compute Hessian from calibration activations.

H = (1/n) * X^T * X + damp * I

Arguments

• activations - Calibration activations [num_samples, in_features]

Returns

Hessian matrix [in_features, in_features]

pub fn compute_hessian_inverse( &self, hessian: &Tensor, ) -> Result<Tensor, PruningError>

Compute Hessian inverse using Cholesky decomposition.

Arguments

• hessian - Symmetric positive definite Hessian matrix

Returns

Inverse of Hessian matrix

Toyota Way: Jidoka

Validates positive definiteness and numerical stability.

pub fn compute_saliency( &self, weights: &Tensor, hessian_inv: &Tensor, ) -> Result<Tensor, PruningError>

Compute saliency scores from weights and Hessian inverse.

Saliency = w^2 / H^{-1}_{jj}

Lower saliency means the weight can be pruned with less error.

pub fn compute_from_activations( &self, weights: &Tensor, activations: &Tensor, ) -> Result<ImportanceScores, PruningError>

Compute importance from weights and activations.

This is the main entry point for SparseGPT importance scoring.

Trait Implementations

impl Clone for SparseGPTImportance

fn clone(&self) -> SparseGPTImportance

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SparseGPTImportance

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

Formats the value using the given formatter.

impl Importance for SparseGPTImportance

fn compute( &self, module: &dyn Module, context: Option<&CalibrationContext>, ) -> Result<ImportanceScores, PruningError>

Compute importance scores for a module’s parameters.

fn name(&self) -> &'static str

Name of this importance method.

fn requires_calibration(&self) -> bool

Whether this method requires calibration data.