# 0xSero/turboquant

TurboQuant: Near-optimal KV cache quantization for LLM inference (3-bit keys, 2-bit values) with Triton kernels + vLLM integration

Repository: https://github.com/0xSero/turboquant
Canonical: https://ross.abutalabs.com/products/turboquant
Language: Python
License: GPL-3.0
License Family: copyleft
Last push: 2026-03-27T22:30:40+00:00

## Health v2 (maintenance only)
Score: 48/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 74, release rhythm 35, longevity 11
- inputs: {"age_days": 161, "days_push": 159, "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 1739, forks 195 (observed 2026-08-28T04:05:29.993142+00:00)

## What it is
TurboQuant is a Python library implementing near-optimal KV cache quantization for LLM inference, compressing keys to 3-bit and values to 2-bit using Triton kernels. It integrates with vLLM and supports dense and MoE architectures on NVIDIA GPUs.

## Use cases
- reduce KV cache memory usage during LLM inference
- double the max context/token capacity of a vLLM deployment
- serve long-context LLMs on limited GPU VRAM
- quantize KV cache to 3-bit keys and 2-bit values
- speed up prefill and decode throughput for large context windows
- run MoE models with mixed full and linear attention layers more efficiently

## When to choose
- you use vLLM and need to fit longer contexts or more concurrent requests in GPU memory
- you want drop-in KV cache compression with minimal quality loss
- you run inference on NVIDIA GPUs (RTX 3090/5090 class) with dense or MoE models

## When to avoid
- your model uses only linear-attention layers, which TurboQuant cannot compress
- you need a non-vLLM inference engine without integration support
- you require permissive licensing since the project is GPL-3.0
- you need CPU-only or non-CUDA inference

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, gpu-computing, caching, machine-learning
- domain: large-language-models, machine-learning, gpu-computing, performance
- platform: python
- tags: kv-cache-quantization, vllm, triton-kernels, inference-optimization, moe, gpu, linux, docker

## Member repositories
- 0xSero/turboquant (main) score 48

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:29.993142+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 2026-08-30T03:30:12.544914+00:00, confidence not recorded.
  - readme: https://github.com/0xSero/turboquant (fetched 2026-08-28T04:05:29.993142+00:00, sha c86933af7ae9)
- Data as of 2026-08-30T08:39:29.467469+00:00.