skainet-io-gguf/sk.ainet.io.gguf.dequant/DequantOps

DequantOps

Shared dequantization operations for GGUF and SafeTensors weight loading.

All functions are module-visible (internal) and stateless. They convert quantized byte payloads into FP32 float arrays.

Functions

blockInfoFor

fun blockInfoFor(type: GGMLQuantizationType): Pair<Int, Int>

Returns (bytesPerBlock, elemsPerBlock) for a given quantization type. Useful for chunked dequantization on single-threaded platforms (WASM).

bytesToFloatArray

fun bytesToFloatArray(bytes: ByteArray): FloatArray

dequantBF16

fun dequantBF16(raw: List<Any>): FloatArray

dequantBF16FromBytes

fun dequantBF16FromBytes(bytes: ByteArray): FloatArray

dequantF16

fun dequantF16(raw: List<Any>): FloatArray

dequantF16FromBytes

fun dequantF16FromBytes(bytes: ByteArray): FloatArray

dequantFromBytes

fun dequantFromBytes(bytes: ByteArray, tensorType: GGMLQuantizationType, nElems: Int): FloatArray

Dispatch dequantization based on tensor type for byte arrays.

dequantFromList

fun dequantFromList(raw: List<Any>, tensorType: GGMLQuantizationType, nElems: Int): FloatArray

Dispatch dequantization based on tensor type for List (in-memory GGUFReader).

dequantIQ4NL

fun dequantIQ4NL(raw: List<Any>, nElems: Int): FloatArray

dequantIQ4XS

fun dequantIQ4XS(raw: List<Any>, nElems: Int): FloatArray

dequantQ2K

fun dequantQ2K(raw: List<Any>, nElems: Int): FloatArray

dequantQ3K

fun dequantQ3K(raw: List<Any>, nElems: Int): FloatArray

dequantQ4_0

fun dequantQ4_0(raw: List<Any>, nElems: Int): FloatArray

dequantQ4_1

fun dequantQ4_1(raw: List<Any>, nElems: Int): FloatArray

dequantQ4K

fun dequantQ4K(raw: List<Any>, nElems: Int): FloatArray

dequantQ5_0

fun dequantQ5_0(raw: List<Any>, nElems: Int): FloatArray

dequantQ5_1

fun dequantQ5_1(raw: List<Any>, nElems: Int): FloatArray

dequantQ5K

fun dequantQ5K(raw: List<Any>, nElems: Int): FloatArray

dequantQ6K

fun dequantQ6K(raw: List<Any>, nElems: Int): FloatArray

dequantQ8_0

fun dequantQ8_0(raw: List<Any>, nElems: Int): FloatArray

dequantQ8_1

fun dequantQ8_1(raw: List<Any>, nElems: Int): FloatArray

dequantQ8K

fun dequantQ8K(raw: List<Any>, nElems: Int): FloatArray

dequantTQ1_0

fun dequantTQ1_0(raw: List<Any>, nElems: Int): FloatArray

dequantTQ2_0

fun dequantTQ2_0(raw: List<Any>, nElems: Int): FloatArray

halfToFloat

fun halfToFloat(hbits: Int): Float

toByteArray

fun toByteArray(raw: List<Any>, tensorName: String): ByteArray

transposeColumnMajorToRowMajor

fun transposeColumnMajorToRowMajor(data: FloatArray, rows: Int, cols: Int): FloatArray

Handle column-major to row-major conversion for GGUF tensors. Data layout is unchanged — only the shape interpretation changes.