# scrya-com/rotorquant

KV cache compression via block-diagonal rotation. Beats TurboQuant: better PPL (6.91 vs 7.07), 28% faster decode, 5.3x faster prefill, 44x fewer params. Drop-in llama.cpp integration.

Repository: https://github.com/scrya-com/rotorquant
Canonical: https://ross.abutalabs.com/products/rotorquant
Homepage: https://www.scrya.com/rotorquant
Language: Python
License Family: other
Last push: 2026-04-23T10:51:05+00:00

## Health v2 (maintenance only)
Score: 50/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 78, release rhythm 35, longevity 11
- inputs: {"age_days": 160, "days_push": 132, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1043, forks 89 (observed 2026-08-28T04:03:21.010516+00:00)

## What it is
RotorQuant is a KV cache compression method for LLM inference that replaces full d×d orthogonal rotations with block-diagonal Clifford rotor transforms, integrated as a drop-in quantization scheme in llama.cpp. It claims better perplexity and substantially faster decode/prefill than TurboQuant at the same compression ratio.

## Use cases
- compress kv cache for long-context llm inference
- speed up llama.cpp decode and prefill
- reduce gpu memory for llm serving
- quantize attention k/v vectors at 3-bit
- compare against turboquant baselines

## When to choose
- you serve LLMs via llama.cpp and need lower memory or higher throughput at low bit-widths
- you want KV cache quantization with minimal quality loss (PPL) and fast kernels on CUDA or Metal

## When to avoid
- you need a permissively licensed dependency - the repo has no license
- you need a framework-agnostic solution outside llama.cpp/Triton
- you require independently verified benchmarks - results come from the authors' own paper

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, caching, gpu-computing, benchmarking
- domain: large-language-models, machine-learning, performance, gpu-computing
- platform: python, cpp, windows
- tags: kv-cache, quantization, clifford-algebra, llama-cpp, cuda, metal, inference-optimization, triton, gpu, linux, macos

## Member repositories
- scrya-com/rotorquant (main) score 50

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:21.010516+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:02:31.834007+00:00, confidence not recorded.
  - readme: https://github.com/scrya-com/rotorquant (fetched 2026-08-28T04:03:21.010516+00:00, sha 951a5af98e95)
  - homepage: https://www.scrya.com/rotorquant (fetched 2026-08-29T13:03:56.799537+00:00, sha 1f3f0e78251d)
  - site_page: https://www.scrya.com/about (fetched 2026-08-29T13:03:56.812773+00:00, sha 737638441470)
  - site_page: https://www.scrya.com/pricing (fetched 2026-08-29T13:03:56.809057+00:00, sha 43b9d6850878)
  - site_page: https://www.scrya.com/faq (fetched 2026-08-29T13:03:56.811032+00:00, sha 81cac69447b6)
- Data as of 2026-08-30T08:39:29.467469+00:00.