Module tensor_quantizer 

TensorQuantizer — Multi-precision tensor quantization for model compression.

Provides production-grade quantization for INT8 (symmetric and asymmetric), INT4, FP16, and BF16, with per-channel support, calibration-percentile outlier suppression, and comprehensive MSE error measurement.

Examples

use ipfrs_tensorlogic::tensor_quantizer::{
    TensorQuantizer, QuantizationMode, QuantizerConfig,
};

let config = QuantizerConfig {
    mode: QuantizationMode::Int8Symmetric,
    per_channel: false,
    channel_dim: 0,
    calibration_percentile: 99.9,
};
let quantizer = TensorQuantizer::new(config);
let values = vec![0.5_f64, -0.3, 0.8, -0.1, 1.0, -1.0];
let dims = vec![6];
let qt = quantizer.quantize(&values, &dims).expect("example: should succeed in docs");
let dq = quantizer.dequantize(&qt).expect("example: should succeed in docs");
assert_eq!(dq.values.len(), 6);

Structs

DequantizedTensor

The result of dequantization — approximate reconstruction of the original values.

QuantizedTensor

A quantized representation of a tensor.

QuantizerConfig

Configuration for TensorQuantizer.

QuantizerStats

Accumulated statistics across multiple TensorQuantizer::quantize calls.

TensorQuantizer

Multi-precision tensor quantizer.

Enums

QuantizationMode

Precision target for quantization.

QuantizerError

Errors produced by TensorQuantizer.

Functions

percentile

Compute the p-th percentile of values using the nearest-rank method.