Skip to main content

turboquant_dequantize

mlx_native::turboquant

Function turboquant_dequantize 

Source 
pub fn turboquant_dequantize(
    packed: &[u8],
    norm: f32,
    config: &TurboQuantConfig,
) -> Result<Vec<f32>>
Expand description

Dequantize a TurboQuant-compressed head vector.

Steps:

  1. Unpack indices
  2. Look up centroid values, scale back from N(0,1) domain
  3. Multiply by norm
  4. Apply inverse FWHT (same as forward)

§Arguments

  • packed — packed index bytes
  • norm — the L2 norm stored during quantization
  • config — quantization configuration

§Returns

Reconstructed vector of length config.head_dim.