[ English | δΈζ | ζ₯ζ¬θͺ ]
LLMC is an off-the-shell tool designed for compressing LLM, leveraging state-of-the-art compression algorithms to enhance efficiency and reduce model size without compromising performance.
English doc is here.
Chinese doc is here.
Docker hub is here.
Aliyun docker: registry.cn-hangzhou.aliyuncs.com/yongyang/llmcompression:[tag]
You can download the Docker image that can run llmc with the following command. Users in mainland China are recommended to use Alibaba Cloud Docker.
docker hub
docker pull llmcompression/llmc:pure-latest
aliyun docker
docker pull registry.cn-hangzhou.aliyuncs.com/yongyang/llmcompression:pure-latest
Community:
-
May 12, 2025: π₯ We now fully support quantization for the
Wan2.1series of video generation models and provide export of truly quantized INT8/FP8 weights, compatible with the lightx2v inference framework. For details, please refer to the lightx2v documentation. -
Feb 7, 2025: π₯ We now fully support quantization of large-scale
MOEmodels likeDeepSeekv3,DeepSeek-R1, andDeepSeek-R1-zerowith671Bparameters. You can now directly load FP8 weights without any extra conversion. AWQ and RTN quantization can run on a single 80GB GPU, and we also support the export of true quantized INT4/INT8 weights. -
Nov 20, 2024: π₯ We now fully support the quantization of β¨
DeepSeekv2(2.5)and otherMOEmodels, as well as β¨Qwen2VL,Llama3.2, and otherVLMmodels. Supported quantization methods include β integer quantization, β floating-point quantization, and advanced algorithms like β AWQ, β GPTQ, β SmoothQuant, and β Quarot. -
Nov 12, 2024: π₯ We have added support for π₯
static per-tensor activation quantizationacross various models and algorithms, covering β integer quantization and β floating-point quantization to further optimize performance and efficiency. Additionally, we now support exporting β¨real quantized modelsand using the VLLM and SGLang backends for inference acceleration. For more details, refer to the VLLM documentation and SGLang documentation. -
Sep 26, 2024: π₯ We now support exporting π₯
FP8 quantized(E4M3, E5M2)models from πLLMCto advanced inference backends such as VLLM and SGLang. For detailed usage, please refer to the VLLM documentation and SGLang documentation. -
Sep 24, 2024: π₯ We have officially released β INT4 and β INT8 models of β¨
Llama-3.1-405B, quantized using πLLMCinsave_lightllmmode. You can download the model parameters here. -
Sep 23, 2024: π₯ We now support exporting β¨
real quantized(INT4, INT8)models from πLLMCto advanced inference backends such as VLLM, SGLang, AutoAWQ, and MLC-LLM for quantized inference deployment, enabling β¨reduced memory usageand β¨faster inference speeds. For detailed usage, please refer to the VLLM documentation, SGLang documentation, AutoAWQ documentation, and MLC-LLM documentation. -
Sep 9, 2024: π₯ We provide some configs of our best practice towards superior performance (see Best Practice here).
-
Sep 3, 2024: π₯ We support opencompass π€ to eval π
LLMCmodel. Follow this doc and have a try! -
Aug 22, 2024: π₯We support lots of small language models, including current SOTA SmolLM(see Supported Model List).
-
Aug 22, 2024: π₯ Additionally, we also support down stream task evaluation through our modified lm-evaluation-harness π€. Specifically, people can first employ
save_transmode(seesavepart in Configuration) to save a weight modified model. After obtaining the transformed model, they can directly evaluate the quantized model referring to run_lm_eval.sh. More details can be found in here. -
Jul 23, 2024: πΊπΊπΊ We release a brand new version benchmark paper:
LLMC: Benchmarking Large Language Model Quantization with a Versatile Compression Toolkit.
Ruihao Gong*, Yang Yong*, Shiqiao Gu*, Yushi Huang*, Chengtao Lv, Yunchen Zhang, Xianglong Liuπ§, Dacheng Tao
(* denotes equal contribution, π§ denotes corresponding author.)
Previous News
-
Jul 16, 2024: π₯We support Wanda/Naive(Magnitude) for llm sparsification and layer-wise mix bits quantization now!
-
Jul 14, 2024: π₯We support rotation based quantization QuaRot now!
-
May 17, 2024: π We support some advanced large models, e.g., LLaVA, Mixtral, LLaMA V3 and Qwen V2 now. Have a try!
-
May 13, 2024: πΊπΊπΊ We release our quantization benchmark paper:
Ruihao Gong*, Yang Yong*, Shiqiao Gu*, Yushi Huang*, Yunchen Zhang, Xianglong Liuπ§, Dacheng Tao
(* denotes equal contribution, π§ denotes corresponding author.)
We modularly and fairly benchmark the quantization techniques considering calibration cost, inference efficiency, and quantized accuracy. Near 600 experiments on diverse models and datasets provide three insightful takeaways on the calibration data, algorithm pipeline, and quantization configuration selection. Based on the takeaways, a best practice for the LLM PTQ pipeline is designed, to achieve the best accuracy and efficiency performance balance under various scenarios.
-
Mar 7, 2024: π We release the quantization part of a powerful and efficient LLM compression tool. Notably, our benchmark paper is coming soonπ.
-
π₯Comprehensive Algorithm Support: Provides a broad range of β¨
SOTA compression algorithms, including β quantization, β mixed-precision quantization, and β sparsity, while maintaining accuracy consistent with the original repositories. β¨Quantization best practices(see πBest Practiceshere) are also available to ensure optimal performance and efficiency. -
π₯Supported Formats: Supports both β¨
quantization(integer and floating-point) and β¨sparsity, specifically including β weight-activation, β weight-only, β mixed-precision quantization, as well as β structured and β unstructured sparsity. -
π₯Wide Model Support: Offers support for a diverse array of β¨
LLM models, including β LLama, β Mistral, β InternLM2, β Qwen2, among others, as well as β MOE(DeepSeekv2, Deepseek-R1) and β VLM(Llama3.2-vision, Qwen2-vl) models (see Supported Model List). -
π₯Multi-backend Compatibility: Seamlessly integrates with various backends for enhanced deployment flexibility. Multiple quantization settings and model formats are compatible with a wide range of backends and hardware platforms, such as β VLLM, β Sglang, β LightLLM, β MLC-LLM, and β AutoAWQ, making it highly versatile(see Section
Backendhere). -
π₯Performance Efficiency: Enables quantization of large LLMs, such as β¨
Llama3.1-405Band β¨DeepSeek-R1-671B, with PPL evaluation on asingle A100/H100/H800 GPU.
Please refer to the πQuick Start section in the documentation.
β BLOOM
β LLaMA
β LLaMA V2
β StarCoder
β OPT
β Falcon
β InternLM2
β Mistral
β LLaMA V3
β Mixtral
β Qwen V2
β LLaVA
β InternLM2.5
β StableLM
β Gemma2
β Phi2
β Phi 1.5
β MiniCPM
β SmolLM
β DeepSeekv2.5
β Qwen MOE
β Qwen2-VL
β InternVL2
You can add your own model type referring to files under llmc/models/*.py.
β VLLM
β LightLLM
β Sglang
β MLC-LLM
β AutoAWQ
β Naive
β AWQ
β GPTQ
β SmoothQuant
β OS+
β OmniQuant
β NormTweaking
β AdaDim
β QUIK
β SpQR
β DGQ
β OWQ
β LLM.int8()
β HQQ
β QuaRot
β SpinQuant (See this branch)
β TesseraQ
β Naive(Magnitude)
β Wanda
β ShortGPT
We develop our code referring to the following repos:
- https://github.com/mit-han-lab/llm-awq
- https://github.com/mit-han-lab/smoothquant
- https://github.com/OpenGVLab/OmniQuant
- https://github.com/IST-DASLab/gptq
- https://github.com/ModelTC/Outlier_Suppression_Plus
- https://github.com/IST-DASLab/QUIK
- https://github.com/Vahe1994/SpQR
- https://github.com/ilur98/DGQ
- https://github.com/xvyaward/owq
- https://github.com/TimDettmers/bitsandbytes
- https://github.com/mobiusml/hqq
- https://github.com/spcl/QuaRot
- https://github.com/locuslab/wanda
- https://github.com/EleutherAI/lm-evaluation-harness
- https://github.com/facebookresearch/SpinQuant
- https://github.com/Intelligent-Computing-Lab-Yale/TesseraQ
If you find our LLM-QBench paper/llmc toolkit useful or relevant to your research, please kindly cite our paper:
@misc{llmc,
author = {llmc contributors},
title = {llmc: Towards Accurate and Efficient LLM Compression},
year = {2024},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/ModelTC/llmc}},
}
@misc{gong2024llmqbench,
title={LLM-QBench: A Benchmark Towards the Best Practice for Post-training Quantization of Large Language Models},
author={Ruihao Gong and Yang Yong and Shiqiao Gu and Yushi Huang and Yunchen Zhang and Xianglong Liu and Dacheng Tao},
year={2024},
eprint={2405.06001},
archivePrefix={arXiv},
primaryClass={cs.LG}
}
@misc{gong2024llmcbenchmarkinglargelanguage,
title={LLMC: Benchmarking Large Language Model Quantization with a Versatile Compression Toolkit},
author={Ruihao Gong and Yang Yong and Shiqiao Gu and Yushi Huang and Chentao Lv and Yunchen Zhang and Xianglong Liu and Dacheng Tao},
year={2024},
eprint={2405.06001},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2405.06001},
}