TurboQuantPresets

Named preset configurations for TurboQuant KV-cache compression.

Presets reflect the practical observation that key precision is often more quality-sensitive than value precision.

Available presets:

Properties

List all available preset names.

fun balanced(numLayers: Int, numHeads: Int, headDim: Int, maxSeqLen: Int): TurboQuantPreset

Balanced preset: TurboQuant-4 for both keys and values.

fun experimentalMax(numLayers: Int, numHeads: Int, headDim: Int, maxSeqLen: Int): TurboQuantPreset

Experimental maximum compression: TurboQuant-3 for both K and V.

fun forModel(preset: String, numLayers: Int, numHeads: Int, headDim: Int, maxSeqLen: Int): TurboQuantPreset

Look up a preset by name and apply model dimensions.

fun safeLowbit(numLayers: Int, numHeads: Int, headDim: Int, maxSeqLen: Int): TurboQuantPreset

Safe low-bit preset: Q8_0 for keys, TurboQuant-4 for values.