Tsinghua University

3 月 14 日消息，清华大学高性能计算研究所翟季冬教授团队、清华系科创企业清程极智今日联合宣布，大模型推理引擎“赤兔 Chitu”现已开源。 据介绍，该引擎首次实现在非英伟达 Hopper 架构 GPU 及各类国产芯片上原生运行 FP8 精度模型，实现 DeepSeek 推理成本降低一半、性能翻番。其定位为“生产级大模型推理引擎”，提供如下特性： 多元算力适配：不仅支持 NVIDIA 最新旗舰到…

Recently, Tsinghua University and Harbin Institute of Technology jointly published a paper that successfully compressed a large model to 1 bit while maintaining the performance of 83%. This achievement marks a major breakthrough in the field of quantization models. In the past, quantization below 2 bits has always been an insurmountable obstacle for researchers, and this attempt at 1-bit quantization has attracted widespread attention from the academic community at home and abroad. The OneBit method proposed in this study is the first attempt to compress a pre-trained large model to a true 1 bit. Through a new 1-bit layer structure, SVID-based parameter initialization and quantization…

Recently, a series of open-source visual models with excellent performance have emerged under the promotion of China's top universities such as Tsinghua University and Zhejiang University, which are open-source alternatives to GPT-4V. Among them, LLaVA, CogAgent and BakLLaVA are three open-source visual language models that have attracted much attention. LLaVA is a large multimodal model trained end-to-end that combines the visual encoder and Vicuna for general vision and language understanding, and has impressive chat capabilities. CogAgent is an open-source visual language model improved on CogVLM, with 11 billion...