QuantizedMatrix

classin packageklyn.math.nn.quantization

public class QuantizedMatrix

QuantizedMatrix

A packed Klyn Q8 matrix with row-local FP32 scales. Loading allocates only the INT8 weights and their scales, never a full floating copy. Explicit CPU primitives consume that representation directly; this class does not claim GPU execution or implement a complete language model.

Example

matrix = QuantizedMatrix(Path("model-q8.safetensors"), "projection.weight")

Properties

Modifier and Type	Member	Description
public readonly property	columns `columns as Long`	Input columns in the original weight matrix.
public readonly property	rows `rows as Long`	Output rows in the original weight matrix.
public readonly property	storageBytes `storageBytes as ULong`	Physical weight and scale bytes, excluding object headers.

Constructors

Modifier and Type	Member	Description
public	QuantizedMatrix `QuantizedMatrix(path as Path, name as String)`	Loads and validates one rank-two matrix from a Klyn Q8 shard.

Methods

Modifier and Type	Member	Description
public	embeddingCPU `embeddingCPU(notnull indices as IList<Long>) as NDArray<Float>`	Decodes selected embedding rows on the CPU, not the entire table.
public	multiplyCPU `multiplyCPU(inputs as NDArray<Float>) as NDArray<Float>`	Computes inputs times this weight matrix's transpose on the CPU.