monatq

Monakhov Tensor Quantiles - approximate quantile tracking for tensors using T-Digest.

monatq maintains one T-Digest per element position across many tensor samples, enabling fast approximate quantile queries (median, percentiles, etc.) over the observed value distribution at each position. Updates are parallelised element-wise via Rayon.

Use Cases

Model interpretability - feed activation or weight tensors through a forward pass and query per-position quantiles to understand how individual neurons or channels behave across inputs.
Quantization & pruning - characterise the value distribution at each position to inform bit-width selection, clipping ranges, or sparsity thresholds without storing all observations in memory.

monatq visualizer

Python

See monatq-py/README.md for the Python bindings, including installation and usage with NumPy and PyTorch.

Rust

cargo add monatq

Usage

use monatq::TensorDigest;

// Track a [3, 4] tensor (12 elements)
let mut digest = TensorDigest::new(&[3, 4], 100);

// Feed samples (row-major flat slices)
for sample in my_tensor_samples {
    digest.update(&sample);
}

// Query the per-element median
let medians: Vec<f32> = digest.quantile(0.5);

// Query multiple quantiles at once
let [p10, p50, p90] = digest.quantiles(&[0.1, 0.5, 0.9])[..] else { panic!() };

// Classify the distribution shape at each position
let distributions = digest.analyze();

Features

Parallel updates - element-wise compression runs in parallel via Rayon
Custom T-Digest - optimised implementation for the tensor case, making per-position quantile tracking practical at tensor scale
Distribution analysis - classify each position as Normal, Uniform, Laplace, or LogNormal by fitting an empirical quantile profile
Save / load - zstd-compressed bincode snapshots via digest.save(path) / TensorDigest::load(path)
Visualisation - built-in HTTP server (digest.visualize()) for browser-based inspection of a tensor

License

Apache-2.0 - see LICENSE.