Q4_KTensorData

Tensor data interface for Q4_K quantized format.

Q4_K block format (256 elements per block, 144 bytes per block):

• 2 bytes: f16 d (main scale)

• 2 bytes: f16 dMin (minimum scale)

• 12 bytes: packed scales (8 sub-blocks × 12 bits each = 96 bits = 12 bytes)

• 128 bytes: 4-bit quantized codes (256 elements / 2 = 128 bytes)

Each sub-block (32 elements):

• 6-bit scale index (0..63)

• 6-bit min index (0..63)

• scale = d * (scaleIdx / 63)

• min = dMin * (minIdx / 63)

Dequantization: outputi = codei * scale + min

Inheritors