polarquant 0.1.0

Rust implementation for Walsh-Hadamard rotation + polar coordinate quantization for LLM weight and KV cache compression
Documentation
  • Coverage
  • 50%
    1 out of 2 items documented0 out of 1 items with examples
  • Size
  • Source code size: 14.82 kB This is the summed size of all the files inside the crates.io package for this release.
  • Documentation size: 1 MB This is the summed size of all files generated by rustdoc for all configured targets
  • Ø build duration
  • this release: 40s Average build duration of successful builds.
  • all releases: 40s Average build duration of successful builds in releases after 2024-10-23.
  • Links
  • Repository
  • crates.io
  • Dependencies
    • —
  • Versions
  • Owners
  • prashanth-up

PolarQuant

Walsh-Hadamard rotation + polar coordinate quantization for LLM weight and KV cache compression.

Overview

PolarQuant compresses neural network weights and KV cache embeddings by:

  1. Block-wise L2 normalization to the unit hypersphere
  2. Walsh-Hadamard rotation to decorrelate values into approximately i.i.d. Gaussian
  3. Recursive polar coordinate transformation
  4. Lloyd-Max optimal codebook quantization of resulting angles

Work in progress. Contributions welcome.