skainet-lang-core/sk.ainet.lang.tensor.ops.turboquant/TurboQuantBlock

TurboQuantBlock

data class TurboQuantBlock(val packedCodes: ByteArray, val scales: FloatArray, val seed: Int, val bits: Int, val elementCount: Int, val residual: EncodedResidual? = null)(source)

A single TurboQuant-encoded block.

Contains all data needed to reconstruct the original float vector.

Constructors

TurboQuantBlock

constructor(packedCodes: ByteArray, scales: FloatArray, seed: Int, bits: Int, elementCount: Int, residual: EncodedResidual? = null)

Properties

Bits per code.

val elementCount: Int

Number of logical float elements.

val isPolarOnly: Boolean

val packedCodes: ByteArray

Bit-packed quantization codes.

val residual: EncodedResidual?

Optional QJL residual (null for PolarOnly).

val scales: FloatArray

Per-group scale factors.

Rotation seed for reproducibility.

val sizeInBytes: Int

Total bytes used by this block.

Functions

open operator override fun equals(other: Any?): Boolean

open override fun hashCode(): Int