Module quantization 

Model quantisation.

Module	Contents
int8	INT8 post-training quantisation, FP16 export, calibration

Supported schemes

Scheme	Memory saving	Quality loss
FP16 (half precision)	2×	Negligible
INT8 symmetric per-tensor	4×	Small (~0.5–1 % on retrieval tasks)
INT8 symmetric per-channel	4×	Very small (~0.1–0.3 %)

Workflow

use sensorlm::quantization::int8::{quantize_model_weights, QuantizedModel};
use std::path::Path;

// Extract weights from a trained model (pseudo-code):
// let weights = extract_linear_weights(&trained_model);
// let qm = quantize_model_weights(config_json, weights.into_iter());
// qm.save(Path::new("model_int8.json")).unwrap();

Modules

int8

INT8 post-training quantisation (PTQ).