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

TurboQuantCodec

object TurboQuantCodec(source)

End-to-end TurboQuant encode/decode codec.

Wires together the full TurboQuant pipeline:

Random rotation (spread quantization error)
Scalar quantization (map to N-bit codes)
Optional QJL residual (preserve inner-product accuracy)
Bit-packing (compact storage)

Supports two variants:

PolarOnly: Steps 1-2-4 (fast, backend-friendly)
PolarPlusQjl: Steps 1-2-3-4 (higher accuracy)

Usage:

val encoded = TurboQuantCodec.encode(vector, config)
val decoded = TurboQuantCodec.decode(encoded)

Functions

fun decode(block: TurboQuantBlock): FloatArray

Decode a TurboQuant block back to float values.

fun encode(input: FloatArray, config: TurboQuantConfig): TurboQuantBlock

Encode a float vector using TurboQuant.

fun encodedSize(elementCount: Int, config: TurboQuantConfig): Int

Compute the byte size of an encoded block.