Struct TurboQuantized 

pub struct TurboQuantized { /* private fields */ }

Available on crate feature alloc only.

Quantized weight vector (owned).

Created by quantize, quantize_weights, quantize_f32, or quantize_i16. Supports inference via predict and serialization via to_bytes.

Weights are stored in Hadamard-rotated space (H * D * w). During prediction, the same rotation is applied to features so the dot product is preserved.

Implementations

impl TurboQuantized

pub fn predict(&self, features: &[f64]) -> f64

Dot product of quantized weights with a feature vector.

Applies the same Hadamard rotation to features, then computes the dot product with the quantized rotated weights over the full padded length. Since HD is orthogonal, w . x == (HD*w) . (HD*x).

pub fn predict_with_scratch(&self, features: &[f64], scratch: &mut [f64]) -> f64

Predict using a caller-provided scratch buffer for the Hadamard rotation.

scratch must have length >= padded_len. This avoids allocation, making it suitable for embedded inference loops.

pub fn dequantize(&self) -> Vec<f64>

Dequantize all weights back to f64 (approximate original space).

Unpacks all padded_len rotated values, then applies the inverse Hadamard rotation to recover approximate original weights.

pub fn n_weights(&self) -> usize

Number of quantized weights.

pub fn padded_len(&self) -> usize

Power-of-2 padded length used for FWHT (needed for scratch allocation).

pub fn mode(&self) -> QuantMode

Quantization mode used.

pub fn compression_ratio(&self) -> f64

Compression ratio vs f64 (original bytes / packed bytes).

pub fn to_bytes(&self) -> Vec<u8>

Serialize to bytes for embedded deployment.

Format (36-byte header):

[n_weights: u32 LE]
[mode: u32 LE]        // 0=Bits8, 1=Bits3_5, 2=Bits2_5
[seed: u64 LE]
[padded_len: u32 LE]
[scale: f64 LE]
[offset: f64 LE]
[packed_words: u32 LE...]