# intel/auto-round

A SOTA quantization toolkit for high-accuracy low-bit LLM inference, seamlessly optimized for CPU/XPU/CUDA, with multi-datatype support and full compatibility with vLLM, SGLang, and Transformers|简洁且高效的量化工具包

Repository: https://github.com/intel/auto-round
Canonical: https://ross.abutalabs.com/products/auto-round
Language: Python
License: Apache-2.0
License Family: permissive
Topics: int4, quantization, rounding, transformers, vllm, mxfp4, nvfp4, gguf, sglang, llms, vlms, diffusers, omni
Last push: 2026-08-26T15:29:11+00:00

## Health v2 (maintenance only)
Score: 91/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 93, longevity 69
- inputs: {"age_days": 972, "days_push": 7, "days_rel": 51, "gap_med": 13, "n_releases_24m": 34}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1588, forks 168 (observed 2026-08-28T04:05:07.919549+00:00)

## What it is
AutoRound is Intel's advanced quantization toolkit for LLMs and vision-language models, using sign-gradient descent to achieve high accuracy at 2-4 bit widths. It supports multiple datatypes and hardware (CPU, XPU, CUDA) and integrates with vLLM, SGLang, and Hugging Face Transformers.

## Use cases
- quantize an LLM to int4 for faster inference
- compress a large language model to run on CPU
- convert models to gguf or mxfp4 formats
- quantize a vision-language model with minimal accuracy loss
- prepare quantized models for vLLM or SGLang deployment
- calibration-free low-bit quantization of open LLMs

## When to choose
- you need high-accuracy 2-4 bit weight-only quantization for LLMs or VLMs
- you deploy on Intel CPU/GPU or NVIDIA CUDA and want vLLM/SGLang compatibility
- you want multiple output formats like GGUF, MXFP4, or NVFP4 from one toolkit

## When to avoid
- you need quantization for non-transformer model architectures
- you want training-time or QAT quantization rather than post-training weight-only
- you require a fully stable, long-term-supported API given rapid experimental changes

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, cli, sdk
- domain: large-language-models, machine-learning, deep-learning, developer-tools
- platform: python, windows
- tags: quantization, int4, low-bit-quantization, signround, vllm, sglang, gguf, mxfp4, nvfp4, vision-language-models, weight-only-quantization, model-compression, linux, macos, gpu, cpu

## Member repositories
- intel/auto-round (main) score 91

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:07.919549+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 2026-08-30T03:55:01.148539+00:00, confidence not recorded.
  - readme: https://github.com/intel/auto-round (fetched 2026-08-28T04:05:07.919549+00:00, sha 1d1f96809c97)
  - registry_pypi: https://pypi.org/pypi/auto-round/json (fetched 2026-08-29T11:25:44.509573+00:00, sha 3e09e6bcd0bf)
- Data as of 2026-08-30T08:39:29.467469+00:00.