Struct ChunkedQuantizedMatrix 

pub struct ChunkedQuantizedMatrix<F> { /* private fields */ }

A memory-mapped quantized matrix that operates on chunks of data to avoid loading the entire matrix into memory at once

Implementations

impl<F> ChunkedQuantizedMatrix<F>

where F: Float + NumAssign + Zero + Sum + One + ScalarOperand + Send + Sync + Debug + FromPrimitive + AsPrimitive<f32> + 'static, f32: AsPrimitive<F>,

pub fn new( matrix: &ArrayView2<'_, F>, bits: u8, method: QuantizationMethod, file_path: &str, ) -> LinalgResult<Self>

Create a new chunked quantized matrix from an explicit matrix

Quantizes the matrix and stores it to disk in chunks to support out-of-core processing.

Arguments

• matrix - Matrix to quantize (will be processed in chunks)
• bits - Bit width for quantization
• method - Quantization method
• file_path - Path to store the quantized data

Returns

A new ChunkedQuantizedMatrix instance

pub fn symmetric(self) -> Self

Mark the operator as symmetric

Returns

Self with the symmetric flag set to true

pub fn positive_definite(self) -> Self

Mark the operator as positive definite

Returns

Self with the positive_definite flag set to true

pub fn params(&self) -> &QuantizationParams

Get the quantization parameters

pub fn apply(&self, x: &ArrayView1<'_, F>) -> LinalgResult<Array1<F>>

Apply the matrix to a vector

This method loads the matrix from disk in chunks and applies each chunk to the input vector, accumulating the results to avoid loading the entire matrix into memory at once.

Arguments

• x - Input vector

Returns

Result vector

pub fn tomatrix_free_op(&self) -> LinalgResult<QuantizedMatrixFreeOp<F>>

Convert to a regular QuantizedMatrixFreeOp

This loads the entire matrix into memory, so it should only be used for matrices that can fit in memory.

Returns

A QuantizedMatrixFreeOp instance

pub fn solve_conjugate_gradient( &self, b: &Array1<F>, max_iter: usize, tol: F, adaptive_precision: bool, ) -> LinalgResult<Array1<F>>

Solve a linear system using out-of-core conjugate gradient

Arguments

• b - Right-hand side vector
• max_iter - Maximum number of iterations
• tol - Convergence tolerance
• adaptive_precision - Whether to adaptively adjust precision during iterations

Returns

Solution vector

pub fn from_file(filepath: &str) -> LinalgResult<Self>

Create a new chunked quantized matrix from an existing file

This loads the metadata from an existing quantized matrix file but doesn’t load the entire matrix into memory.

Arguments

• file_path - Path to the quantized matrix file

Returns

A new ChunkedQuantizedMatrix instance

Trait Implementations

impl<F> MatrixFreeOp<F> for ChunkedQuantizedMatrix<F>

where F: Float + NumAssign + Zero + Sum + One + ScalarOperand + Send + Sync + Debug + 'static + FromPrimitive + AsPrimitive<f32>, f32: AsPrimitive<F>,

fn apply(&self, x: &ArrayView1<'_, F>) -> LinalgResult<Array1<F>>

Apply the linear operator to a vector

fn nrows(&self) -> usize

Get the number of rows in the linear operator

fn ncols(&self) -> usize

Get the number of columns in the linear operator

fn is_symmetric(&self) -> bool

Check if the linear operator is symmetric

fn is_positive_definite(&self) -> bool

Check if the linear operator is positive definite