# tonbistudio/turboquant-pytorch

From-scratch PyTorch implementation of Google's TurboQuant (ICLR 2026) for LLM KV cache compression. 5x compression at 3-bit with 99.5% attention fidelity.

Repository: https://github.com/tonbistudio/turboquant-pytorch
Canonical: https://ross.abutalabs.com/products/turboquant-pytorch
Language: Python
License: MIT
License Family: permissive
Last push: 2026-04-23T19:45:18+00:00

## Health v2 (maintenance only)
Score: 50/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 78, release rhythm 35, longevity 11
- inputs: {"age_days": 161, "days_push": 132, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, young
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1040, forks 138 (observed 2026-08-28T04:03:20.117644+00:00)

## What it is
A from-scratch PyTorch implementation of Google's TurboQuant (ICLR 2026) algorithm for compressing LLM key-value caches, including an improved V3 variant that removes the paper's QJL component. It provides quantized KV cache utilities with verified compression/attention-fidelity benchmarks.

## Use cases
- compress llm kv cache to fit longer contexts in gpu memory
- quantize key-value cache to 3-4 bits in pytorch
- reproduce turboquant paper results
- run long-context llm inference with lower memory usage
- benchmark kv cache compression against attention fidelity
- implement vector quantization for transformer caches

## When to choose
- you need KV cache compression for LLM inference in PyTorch
- you want to reproduce or extend the TurboQuant paper
- you need verified, benchmarked quantization with honest results reporting

## When to avoid
- you need production-ready, drop-in KV cache compression for a serving stack
- you expect the advertised 5x compression at 3-bit to work for generation - tests show 3-bit output is broken
- you use non-NVIDIA GPUs or non-PyTorch frameworks

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, llm-inference, caching, compression
- domain: large-language-models, deep-learning, machine-learning, performance
- platform: python, windows, cross-platform
- tags: kv-cache-compression, vector-quantization, pytorch, quantization, llm-optimization, research-reproduction, gpu, linux

## Member repositories
- tonbistudio/turboquant-pytorch (main) score 50

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:20.117644+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 2026-08-30T07:03:32.575685+00:00, confidence not recorded.
  - readme: https://github.com/tonbistudio/turboquant-pytorch (fetched 2026-08-28T04:03:20.117644+00:00, sha 3e0cea7b84ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.