TurboQuantPolarQjl
data class TurboQuantPolarQjl(val bitsPerElement: Int = 4, val residualBits: Int = 1, val blockSize: Int = 128) : TensorEncoding(source)
TurboQuant PolarPlusQjl encoding: rotation + scalar quantization + QJL residual + bit-packing.
Closest to the official TurboQuant paper. The QJL residual stage preserves inner-product accuracy at the cost of additional storage.
Parameters
bitsPerElement
Bits for the primary quantization (2, 3, 4, or 8)
residualBits
Bits for the QJL residual (typically 1 or 2)
blockSize
Elements per block
Functions
Link copied to clipboard
Physical bytes required to store elementCount logical elements in this encoding, or null if the encoding is opaque/variable.