2022级 刘一茳


刘一茳

 

个人介绍:

刘一茳,本科毕业于西安电子科技大学,研究生毕业于英国爱丁堡大学,南京大学电子学院博士研究生,主要研究方向为多模态大模型的效率优化,探索模型量化、剪枝、高效调参等技术,提升模型训练及推理的能效,研究工作发表在CVPR/ICCV等国际会议。


研究课题:

神经网络量化加速中量化感知训练和训练后量化的研究;

去噪扩散概率模型的应用和量化加速;

多模态大模型的高效微调技术。


研究成果:

(CVPR '24) CDCCA, a Self-Corrected Multimodal Large Language Model designed to optimize the performance of models deployed on client devices by leveraging advanced cloud capabilities.

[Paper] Cloud-Device Collaborative Learning for Multimodal Large Language Models

[Code] https://github.com/2644521362/Cdcca


(CVPR '24)  In this paper, we introduce PromptCoT, an innovative enhancer that autonomously refines prompts for users. 

[Paper] PromptCoT: Align Prompt Distribution via Adapted Chain of Thought

[Code] https://github.com/SanGibb/PromptCoT



(ICCV '23) Compress diffusion models to accelerate the generation process through post-training quantization, and propose time-step-aware calibration scheme to deal with the changing output distributions in diffusion models over time steps:

[Paper] Q-Diffusion: Quantizing Diffusion Models; https://arxiv.org/abs/2302.04304

[Website] https://xiuyuli.com/qdiffusion

IMG_256


 

(CVPR '23) Propose NoisyQuant, a quantizer-agnostic enhancement for the posttraining activation quantization performance of vision transformers by actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer:

[Paper] NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers; https://arxiv.org/abs/2211.16056

1672715942316


 

(开源项目) GitHub开源仓库LLaMA2-AccessoryStars 1.5k+)核心开发者,联合上海人工智能实验室打造多模态大模型的预训练、微调、量化和测评等完整工具链:

[Website] https://github.com/Alpha-VLLM/LLaMA2-Accessory

IMG_256


 

E-mail:liuyijiang@smail.nju.edu.cn