VQ-bench: A Composable Vector Quantization Framework

Vector quantization is an old problem that has recently become central to AI infrastructure, sparking a surge of engineering and research activity. This paper responds to that surge by proposing a unified framework for developing and benchmarking new quantization algorithms.

The framework identifies 7 common conceptual quantization primitives and shows how to compose them arbitrarily, then re-expresses 25 common quantizers as pipelines of these primitives, enabling systematic comparison and modular design.

As a concrete contribution, the authors publish VQ-bench as open-source for further extension and provide reproducible benchmarks publicly, aiming to standardize evaluation in this rapidly evolving area.