🚀 AutoRound Meets SGLang: Enabling Quantized Model Inference with AutoRound

by: By Intel Neural Compressor Team, Nov 14, 2025

Overview

We are thrilled to announce an official collaboration between SGLang and AutoRound, enabling low-bit quantization for efficient LLM inference.

Through this integration, developers can now quantize large models with AutoRound’s signed-gradient optimization and directly deploy them in SGLang’s efficient runtime, achieving low-bit model inference with minimal accuracy loss and significant latency reduction.

What Is AutoRound?

AutoRound is an advanced post-training quantization (PTQ) toolkit designed for Large Language Models (LLMs) and Vision-Language Models (VLMs). It uses signed gradient descent to jointly optimize weight rounding and clipping ranges, enabling accurate low-bit quantization (e.g., INT2 - INT8) with minimal accuracy loss in most scenarios. For example, at INT2 precision, it outperforms popular baselines by up to 2.1x higher in relative accuracy. At INT4 precision, AutoRound continues to hold a competitive edge in most cases. The image below provides an overview of the core algorithm in AutoRound.

Full technical details are presented in the AutoRound paper:

👉 Optimize Weight Rounding via Signed Gradient Descent for the Quantization of LLMs

AutoRound algorithm overview

Despite its robust performance, AutoRound remains fast and lightweight—quantizing a 72B model takes only 37 minutes on a single GPU under light mode.

It further supports mixed-bit tuning, lm-head quantization, GPTQ/AWQ/GGUF format exports, and customizable tuning recipes.

AutoRound Highlights

AutoRound is not only focused on algorithmic innovation and exploration, but also widely recognized for its completeness in quantization engineering.

Accuracy: deliver superior accuracy at low-bit precision

Average accuracy of 10+ tasks at INT4 weight

Schemes: support weight-only quantization, weight & activation quantization, dynamic and static for activation quantization
Mixed-bits: propose an effective algorithm to generate mixed-bits / other data types schemes in minutes
Broad Compatibility:
- Support nearly all popular LLM architectures and over 10 vision-language models (VLMs)
- Support Devices: CPU, Intel GPU, CUDA
- Support Data Types: INT2 - INT8, MXFP4, NVFP4, FP8, and MXFP8
Efficiency: Enables block-wise tuning to lower VRAM usage without sacrificing throughput yet fast

Quantization time cost comparison

Community adoption:
- Work seamlessly with SGLang, TorchAO, Transformers, and vLLM
- Widely used by HuggingFace model hubs such as Intel, OPEA, Kaitchup, and fbaldassarri with approximately two million downloads
Export Formats:
- AutoRound
- GPTQ
- AWQ
- GGUF
- Compressed-tensor (initial support)

Integration Overview

SGLang provides a next-generation inference runtime built for scalable, low-latency LLM deployment. Its multi-modal, multi-GPU, and streaming execution model enables both chat and agentic reasoning workloads with exceptional efficiency.

SGLang’s flexible architecture now offers native hooks for quantized model loading, unlocking AutoRound’s full potential for deployment.

1. Quantize with AutoRound

AutoRound automatically optimizes weight rounding and exports quantized weights that compatible with SGLang.

1.1 API Usage

# for LLM
from auto_round import AutoRound
model_id = "meta-llama/Llama-3.2-1B-Instruct"
quant_path = "Llama-3.2-1B-Instruct-autoround-4bit"
# Scheme examples: "W2A16", "W3A16", "W4A16", "W8A16", "NVFP4", "MXFP4" (no real kernels), "GGUF:Q4_K_M", etc.
scheme = "W4A16"
format = "auto_round"
autoround = AutoRound(model_id, scheme=scheme)
autoround.quantize_and_save(quant_path, format=format) # quantize and save

1.2 CMD Usage

auto-round \
    --model Qwen/Qwen2-VL-2B-Instruct \
    --bits 4 \
    --group_size 128 \
    --format "auto_round" \
    --output_dir ./tmp_autoround

2. Deploying with SGLang

SGLang supports AutoRound-quantized models directly (Version>=v0.5.4.post2). It is compatible with SGLang-supported modeling architectures, including common LLM, VLM, and MoE models, and also supports inference and evaluation of AutoRound mixed-bit quantized models.

2.1 OpenAI-Compatible Inference Usage

from sglang.test.doc_patch import launch_server_cmd
from sglang.utils import wait_for_server, print_highlight, terminate_process

# This is equivalent to running the following command in your terminal
# python3 -m sglang.launch_server --model-path Intel/DeepSeek-R1-0528-Qwen3-8B-int4-AutoRound --host 0.0.0.0

server_process, port = launch_server_cmd(
    """
python3 -m sglang.launch_server --model-path Intel/DeepSeek-R1-0528-Qwen3-8B-int4-AutoRound \
 --host 0.0.0.0 --log-level warning
"""
)
wait_for_server(f"http://localhost:{port}")

2.2 Offline Engine API Inference Usage

import sglang as sgl

llm = sgl.Engine(model_path="Intel/DeepSeek-R1-0528-Qwen3-8B-int4-AutoRound")

prompts = ["Hello, my name is"]
sampling_params = {"temperature": 0.6, "top_p": 0.95}

outputs = llm.generate(prompts, sampling_params)
for prompt, output in zip(prompts, outputs):
    print(f"Prompt: {prompt}\nGenerated text: {output['text']}")

More flexible configurations and deployment options are waiting for you to explore!

Quantization Roadmap

AutoRound’s quantization benchmark results demonstrate robust accuracy retention at low precision. The results below highlight AutoRound’s strong advantages and potential in MXFP4, NVFP4, and mixed-bits model quantization. Note that the accuracy result is measured by average accuracy across lambada_openai, hellaswag, piqa, winogrande, and mmlu task.

As part of AutoRound roadmap, we plan to continue enhancing MXFP4 & NVFP4 accuracy for common models and auto mixed-bits quantization in the upcoming releases.

MXFP4 & NVFP4 Quantization. RTN (Round-to-nearest) algorithm is baseline, and 'alg_ext' option indicates experimental optimization algorithms enabled.

mxfp4 llama3.1-8B-Instruct Qwen2-7.5-Instruct Phi4 Qwen3-32B

RTN 0.6212 0.6550 0.7167 0.6901

AutoRound 0.6686 0.6758 0.7247 0.7211

AutoRound+alg_ext 0.6732 0.6809 0.7225 0.7201

nvfp4 llama3.1-8B-Instruct Qwen2-7.5-Instruct Phi4 Qwen3-32B

RTN 0.6876 0.6906 0.7296 0.7164

AutoRound 0.6918 0.6973 0.7306 0.7306

AutoRound+alg_ext 0.6965 0.6989 0.7318 0.7295

mxfp4	llama3.1-8B-Instruct	Qwen2-7.5-Instruct	Phi4	Qwen3-32B
RTN	0.6212	0.6550	0.7167	0.6901
AutoRound	0.6686	0.6758	0.7247	0.7211
AutoRound+alg_ext	0.6732	0.6809	0.7225	0.7201

nvfp4	llama3.1-8B-Instruct	Qwen2-7.5-Instruct	Phi4	Qwen3-32B
RTN	0.6876	0.6906	0.7296	0.7164
AutoRound	0.6918	0.6973	0.7306	0.7306
AutoRound+alg_ext	0.6965	0.6989	0.7318	0.7295

Auto MXFP4 & MXFP8 Mixed-Bits Quantization

Average bits	Llama3.1-8B-I	Qwen2.5-7B-I	Qwen3-8B	Qwen3-32B
BF16	0.7076 (100%)	0.7075 (100%)	0.6764 (100%)	0.7321 (100%)
4-bit	0.6626 (93.6%)	0.6550 (92.6%)	0.6316 (93.4%)	0.6901 (94.3%)
4.5-bit	0.6808 (96.2%)	0.6776 (95.8%)	0.6550 (96.8%)	0.7176 (98.0%)
5-bit	0.6857 (96.9%)	0.6823 (96.4%)	0.6594 (97.5%)	0.7201 (98.3%)
6-bit	0.6975 (98.6%)	0.6970 (98.5%)	0.6716 (99.3%)	0.7303 (99.8%)

Conclusion

The integration of AutoRound and SGLang marks a major milestone in efficient AI model deployment. This collaboration bridges precision optimization and runtime scalability, allowing developers to move seamlessly from quantization to real-time inference with minimal friction. AutoRound’s signed-gradient quantization ensures high fidelity even at extreme compression ratios, while SGLang’s high-throughput inference engine unlocks the full potential of low-bit execution across CPUs, GPUs, and multi-node clusters.

Looking forward, we aim to expand support for advanced quantization formats, optimize kernel efficiency, and bring AutoRound quantization into** **broader multimodal and agentic workloads. Together, AutoRound and SGLang are setting a new standard for intelligent, efficient, and scalable LLM deployment.