Struct BlockSparseGradient 

pub struct BlockSparseGradient {
    pub blocks: Vec<Vec<f32>>,
    pub block_indices: Vec<(usize, usize)>,
    pub block_size: usize,
    pub shape: Vec<usize>,
    pub block_shape: Vec<usize>,
    pub block_sparsity_ratio: f32,
}

Block-sparse gradient for structured sparsity

Fields

blocks: Vec<Vec<f32>>

Non-zero blocks

block_indices: Vec<(usize, usize)>

Block indices (row, col) for 2D or flat index for 1D

block_size: usize

Block size

shape: Vec<usize>

Shape of the original gradient

block_shape: Vec<usize>

Number of blocks in each dimension

block_sparsity_ratio: f32

Sparsity ratio at block level

Implementations

impl BlockSparseGradient

pub fn from_dense( gradient: &[f32], shape: Vec<usize>, block_size: usize, threshold: f32, ) -> Result<Self>

Create block-sparse gradient from dense gradient

Only rank-1 and rank-2 gradients have a block-sparse layout defined here; higher-rank gradients (e.g. 4-D convolution weights) are reported as unsupported so the caller can fall back to a dense update. A sparse-update optimisation must degrade to correct dense behaviour, never abort.

Errors

Returns TorshError::UnsupportedOperation if shape is not rank 1 or 2.

pub fn to_dense(&self) -> Vec<f32>

Convert back to dense gradient

pub fn is_worth_block_sparse(&self, min_sparsity_ratio: f32) -> bool

Check if block representation is beneficial

Trait Implementations

impl Clone for BlockSparseGradient

fn clone(&self) -> BlockSparseGradient

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for BlockSparseGradient

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

Formats the value using the given formatter.