Efficient ML Algorithm
Algorithm and architecture innovations that improve capability while reducing compute, memory, and deployment cost.
通过算法与架构创新提升模型能力,同时降低计算、显存与部署成本。
Future Open Source Research
面向未来的开源研究
We study how to make machine learning more capable, efficient, and practical at scale through four connected directions: algorithm and architecture innovation, systems optimization, quantization, and post-training-driven modeling.
我们研究如何让机器学习在大规模场景下更强大、更高效、更实用,聚焦四个相互关联的方向:算法与架构创新、系统优化、量化,以及后训练驱动的建模。
Welcome to the Future Machine Learning & Systems (FutureMLS) Lab. We work at the intersection of machine learning and computer systems. Today's foundation models are remarkably capable but costly to train and serve. Our mission is to close the gap between rapidly growing model capability and the real-world cost of deploying these models.
欢迎来到 Future Machine Learning & Systems (FutureMLS) 实验室。我们的研究处在机器学习与计算机系统的交叉地带。今天的基础模型能力惊人,但训练与部署的成本同样高昂。我们的使命,是弥合模型能力的快速增长与真实部署成本之间的鸿沟。
We pursue algorithm-system co-design across four connected themes: Efficient ML Algorithm for algorithm and architecture innovation, Efficient ML System for systems optimization, Quantization as a core research focus, and Modeling for improving models through training. Our work spans the AI stack, from methods and model design to kernels, runtimes, and serving systems, and is open-source, reproducible, and built to be used.
我们围绕四个相互关联的主题推进算法—系统协同设计:Efficient ML Algorithm 负责算法与架构创新,Efficient ML System 负责系统层面的优化,Quantization 是我们的核心研究方向,Modeling 则通过训练让模型变得更强。我们的工作贯穿整个 AI 技术栈——从方法与模型设计,到算子、运行时与服务系统——并且坚持开源、可复现、可直接使用。
Zhongzhu Zhou (Charlie Zhou) is the founder and principal investigator of the Future Machine Learning & Systems Lab. He is a Senior Research Scientist on the Turbo Team at Together AI, and earned his Ph.D. at the School of Computer Science, University of Sydney.
周中柱(Charlie Zhou)是 Future Machine Learning & Systems 实验室的创始人与首席研究员。他现任 Together AI Turbo 团队高级研究科学家,并于悉尼大学计算机科学学院取得博士学位。
His research spans efficient machine learning and systems — from pretraining quality to efficient algorithms and algorithm–system co-design that bridges emerging ML/LLM methods and real-world applications, improving both productivity (usable, robust stacks) and performance (throughput, memory, and cost-efficiency). He received his B.Eng. (Hons) from Sun Yat-sen University, and has interned at Dolby, the DeepSpeed team at Microsoft, and Tencent.
他的研究覆盖高效机器学习与系统——从预训练质量,到高效算法,再到连接前沿 ML/LLM 方法与真实应用的算法—系统协同设计,同时提升研发效率(好用且稳健的技术栈)与运行性能(吞吐、显存与成本)。他本科毕业于中山大学并获荣誉学位,曾在 Dolby、微软 DeepSpeed 团队与腾讯实习。
He leads projects across the lab's four themes, including OSCAR (2-bit KV-cache quantization) and CARE (covariance-aware Multi-Head Latent Attention).
他主导实验室四个方向上的多个项目,包括 OSCAR(2-bit KV-cache 量化)与 CARE(协方差感知的 Multi-Head Latent Attention)。
Four directions, one goal: efficient and capable AI at scale.
四个方向,一个目标:让大规模 AI 既高效又强大。
Algorithm and architecture innovations that improve capability while reducing compute, memory, and deployment cost.
通过算法与架构创新提升模型能力,同时降低计算、显存与部署成本。
System-level optimizations that make efficient methods practical end-to-end, from kernels and runtimes to high-throughput serving.
从算子、运行时到高吞吐服务,用系统层面的优化让高效方法端到端地落地。
A core research focus on low-bit weight, activation, and KV-cache quantization that preserves accuracy while cutting memory and compute.
我们的核心方向:低比特的权重、激活与 KV-cache 量化,在压缩显存与计算的同时保持精度。
Model improvement through training optimization, architecture design, and adaptation methods that make models stronger and easier to use.
通过训练优化、架构设计与适配方法改进模型,让模型更强、也更易用。
Happy to see community interest in bringing OSCAR toward vLLM serving workflows. Grateful to the contributors pushing broader support for deployable rotation-based KV-cache quantization in mainstream LLM serving stacks.
很高兴看到社区正在把 OSCAR 推向 vLLM 的服务链路。感谢各位贡献者,让基于旋转的可部署 KV-cache 量化在主流 LLM 服务栈中获得更广泛的支持。
OSCAR has been covered by multiple tech media outlets, including Towards AI, ModelScope, MarkTechPost, QbitAI (量子位), and Synced (机器之心), reaching 100,000+ reads across platforms. Grateful for the community interest in deployable 2-bit KV-cache quantization.
OSCAR 被多家科技媒体报道,包括 Towards AI、ModelScope、MarkTechPost、量子位与机器之心,全平台阅读量超过 100,000 次。感谢社区对可部署 2-bit KV-cache 量化的关注。
Released Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation. The codebase distills Qwen/Llama softmax-attention Transformers into hybrid linear-attention students built on GatedDeltaNet, using Taylor-series-informed initialization before staged distillation.
发布 Taylor-Calibrate:面向混合线性注意力蒸馏的原理化初始化方法。该代码库把 Qwen/Llama 的 softmax attention Transformer 蒸馏为基于 GatedDeltaNet 的混合线性注意力学生模型,并在分阶段蒸馏前先完成基于泰勒展开的初始化。
Our paper OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization has been accepted by ACL SELVA 2026!
我们的论文 OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization 已被 ACL SELVA 2026 接收!
Huge thanks to the open-source community — OSCAR has reached 500 stars on GitHub and now supports local llama.cpp usage, making 2-bit KV-cache quantization easier to try on local LLM deployments.
衷心感谢开源社区——OSCAR 的 GitHub star 数已达 500,并且现已支持在本地 llama.cpp 中使用,让 2-bit KV-cache 量化更容易在本地 LLM 部署上试用。
OSCAR crosses 300 ★ on GitHub in its first week — thanks to the open-source community.
OSCAR 上线首周 GitHub star 数突破 300 ★——感谢开源社区。
OSCAR released — 2-bit KV-cache serving at 2.28 effective bits/element with near-BF16 accuracy on Qwen3 and GLM-4.7.
OSCAR 正式发布——在 Qwen3 与 GLM-4.7 上以 2.28 等效 bit/元素实现 2-bit KV-cache 服务,精度接近 BF16。
CARE presented at ICLR 2026: covariance-aware, rank-enhanced decomposition for Multi-Head Latent Attention.
CARE 在 ICLR 2026 发表:面向 Multi-Head Latent Attention 的协方差感知、秩增强分解方法。
A small group of researchers and advisors building in the open.
一支小而专注的研究者与顾问团队,坚持开放地做研究。
Advisor
顾问
Principal Research scientist in efficient ML and low-bit quantization, with extensive work across the DeepSpeed and Together AI.
高效机器学习与低比特量化方向的首席研究科学家,在 DeepSpeed 与 Together AI 有大量相关工作。
Advisor
顾问
Professor at the University of Sydney; high-performance computing and ML systems.
悉尼大学教授,研究方向为高性能计算与机器学习系统。
Ph.D. Student, UT Austin
博士生,德克萨斯大学奥斯汀分校
Quantization-aware training for efficient and accurate large language models.
面向高效且高精度大语言模型的量化感知训练。
Student Researcher
学生研究员
Post-training quantization and KV-cache compression for efficient LLM serving; contributor to OSCAR.
面向高效 LLM 服务的训练后量化与 KV-cache 压缩;OSCAR 项目贡献者。
Selected projects from the lab — open the preview page for details, papers, and code.
实验室的精选项目——点击进入可查看详情、论文与代码。
2-bit KV-cache quantization · 2026
Attention-aware offline rotations compress the KV cache to 2.28 bits/element — ~8× memory reduction and up to ~7× higher throughput with near-BF16 accuracy.
View project → 查看项目 →
Multi-Head Latent Attention · ICLR 2026
Covariance-aware low-rank decomposition that converts pretrained GQA/MHA into MLA — up to 215× lower one-shot perplexity at matched KV budgets.
View project → 查看项目 →
Hybrid Linear Attention Distillation · 2026
Taylor-guided initialization for converting Transformers into hybrid GatedDeltaNet students — up to 88× lower initial perplexity and 4.9–9.2× fewer distillation tokens.
View project → 查看项目 →We welcome collaborators, prospective students, and contributors who care about efficient, open machine learning and systems.
我们欢迎关注高效、开放的机器学习与系统的合作者、有意申请的同学,以及开源贡献者。