Zhongzhu Zhou

vibesys. Can AI Agents Build Bespoke Systems?

Taylor-Calibrate. Taylor-Calibrate: Principled Initialization for Hybrid Linear Attention Distillation

auto-gpu-kernel. Winner 🏆 (Agent-only) MLSys 2026 - FlashInfer AI Kernel Generation Contest for the DeepSeek Sparse Attention (DSA) track with an average speedup of 34.93x

slime. slime is an LLM post-training framework for RL Scaling.

flash-llm. Flash-LLM: Enabling Cost-Effective and Highly-Efficient Large Generative Model Inference with Unstructured Sparsity

fp6_llm. An efficient GPU support for LLM inference with x-bit quantization (e.g. FP6,FP5).

Kitty. Algorithm-System Co-design: accurate and efficient 2-bit KV cache quantization for LLM Inference.

OScaR-KV-Quant. 🏆 OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond — redefining the accuracy-efficiency Pareto front for X-LLMs KV quantization.

OSCAR. OSCAR: Offline Spectral Covariance-Aware Rotation for 2-bit KV Cache Quantization

Loom. You drive Claude Code / Codex — Loom keeps the worktrees, plans, diffs, and notes tidy.

ccblog. 🚀 Production-Ready Multi-Agent System for WeChat Article Generation - Automated pipeline that actually works! | 真正能用的微信公众号AI生成系统：多Agent协同、自动化处理、通俗易懂

parcae. Stable Looped Models and their Scaling Laws

InferenceX. Open Source Continuous Inference Benchmark Research Platform — Kimi K3 2.8T, MiniMax M3, DeepSeekv4, GLM5 - GB200 NVL72 vs MI355X vs B200 vs GB300 NVL72 & soon™ TPUv6e/v7/Trainium2/3 | 开源持续推理基准研究平台 — Kimi K2.7-Code、MiniMax M3、DeepSeekv4、GLM5 - GB200 NVL72 vs MI355X vs B200 vs GB300 NVL72，即将推出™ TPUv6e/v7/Trainium2/3

I-DLM. Python

LLMs-from-scratch. Implement a ChatGPT-like LLM in PyTorch from scratch, step by step

aurora. Python

xorl. XoRL

CARE. CARE: Covariance-Aware and Rank-Enhanced Decomposition for Enabling Multi-Head Latent Attention

CREST. CREST is a training-free test-time steering framework that discovers cognitive heads via simple offline calibration and then rotates activations during decoding to guide the model’s reasoning—preserving norms to avoid per-model hyperparameter tuning. This improves accuracy and reduces tokens across models and datasets.

Awesome-ML-SYS-Tutorial. My learning notes for ML SYS.

ccf-deadlines. ⏰ Agenticly track worldwide conference deadlines (Website, Python Cli, Wechat Applet)

ladder-residual-inference. Python

neural-collapse. Generic library for neural collapse and several derivative works on the phenomenon.

CorDA. CorDA: Context-Oriented Decomposition Adaptation of Large Language Models for task-aware parameter-efficient fine-tuning(NeurIPS 2024)

ddl. backend, web client and admin page for Don't Destroy Life (DDL)

pytorch. Tensors and Dynamic neural networks in Python with strong GPU acceleration

EmReal. A digital twin framework, EmReal, with the emulation of hardware components for pre-training in RL and real-time controlling of a real robot.

Hybrid-Share. C

sysConferences. System conferences crawler and timeline

Pre-Expedite. Pre-Expedite: Use Hierarchical Structure Space for Improving the performance of Accessing Small Files in Parallel File System

Notes. Store notes for my phd research

zzz0906.github.io. Zhongzhu Zhou's personal blog | Using Hexo | Theme by Next

DeepSpeed. DeepSpeed is a deep learning optimization library that makes distributed training and inference easy, efficient, and effective.

DeepSpeedExamples. Example models using DeepSpeed

kube-batch. A batch scheduler of kubernetes for high performance workload, e.g. AI/ML, BigData, HPC

PathPlanning. Common used path planning algorithms with animations.

LeetCode. Documenting my LeetCode training journey during my doctoral studies.

DDLs. An Android app for managing DDLs.

SYSU-Exam. 收集整理SYSU期末考试卷子、资料

SevenSeconds. Java

Paddle. PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署）