skainet-lang-core/sk.ainet.lang.tensor.storage/PackedBlockStorage

PackedBlockStorage

interface PackedBlockStorage(source)

Shared contract for all packed/quantized block tensor storage formats.

Instead of each quantization format (Q4_K, Q8_0, Ternary, …) inventing its own loader, planner, and backend handling path, all packed formats implement this interface. Backends and planners can dispatch on encoding without knowing every possible quantization scheme.

Individual formats still expose format-specific accessors (sub-block scales, code extraction, etc.) through their own sub-interfaces.

Inheritors

Q4_KBlockTensorData

Q8_0BlockTensorData

Ternary2BitTensorData

Properties

abstract val blockCount: Int

Number of blocks in this storage.

abstract val blockSize: Int

Number of logical elements per block.

open val elementCount: Long

Logical element count.

abstract val encoding: TensorEncoding

The physical encoding describing the block layout.

abstract val packedData: ByteArray

Raw packed byte data containing all blocks.

open val physicalBytes: Long

Physical byte size of the packed data.

abstract val shape: Shape

The logical shape of the tensor (element count, not block count).

Functions

dequantizeBlock

abstract fun dequantizeBlock(blockIdx: Int, output: FloatArray, outputOffset: Int = 0)

Dequantize a single block to float values.

open fun toFloatArray(): FloatArray

Dequantize the entire tensor to a FloatArray. Default implementation calls dequantizeBlock for each block.

toTensorStorage

open fun toTensorStorage(logicalType: LogicalDType = LogicalDType.FLOAT32, placement: Placement = Placement.CPU_HEAP): TensorStorage

Convert this packed storage to a TensorStorage descriptor.