Crate quantize 

quantize

A tiny, readable quantization library — block-wise symmetric or asymmetric, any bit width.

Example

use quantize::quantize;

let weights = [0.42_f32, -0.10, 0.70, -0.50];

// 8-bit, block-size-32, f32 scales
let q = quantize::<f32, 8, 32>(&weights).unwrap();
let back = q.dequantize();

assert!((back[0] - weights[0]).abs() < 0.01);

BITS and BLOCK are const generics, so quantize::<f32, 4, 32>(...), quantize::<f32, 8, 64>(...), etc. all compile to specialized code.

See symmetric, asymmetric, and adaptive for the other schemes. To learn how the library got here, please see chapters/.

Modules

adaptive

Mixed-precision: pick bits per block from a tolerance.

asymmetric

Asymmetric quantization: scale and zero-point per group.

error

Recoverable failures from quantization and dequantization.

learned

After codes are chosen, pick a better scale and zero-point.

packed

Bit-packed signed integer codes.

params

Integer grids, scale selection, and mixed-precision bit choice.

scale

Scale trait — how per-block scale factors are stored.

scheme

Runtime scheme selection — one entry point, scheme-specific I/O inside.

symmetric

Symmetric quantization: one scale per group, codes centered on zero.

tensor

One enum, one variant per scheme.

Structs

Packed

Packed signed codes plus the bit width they were written with.

Enums

Error

An error produced by a fallible quantization API.

Quantized

Packed codes and the scheme that produced them.

Scheme

Which algorithm to run. Pick this from config or a CLI flag; the returned Quantized variant is the scheme that ran.

Traits

Scale

A type that can serve as a per-block scale (or zero-point) factor.

Functions

quantize

Quantize values into fixed-size blocks of BLOCK with BITS-wide codes.

quantize_tensor

Quantize the entire tensor with a single scale.

Type Aliases

Result

Result alias for this crate.