Xiuyu Li

hyperframes. Write HTML. Render video. Built for agents.

DeepSpec. DeepSpec: a full-stack codebase for training and evaluating speculative decoding algorithms

VeloQ. Agent-friendly GPU profile-query CLI

kernel-design-agents.

codex. Lightweight coding agent that runs in your terminal

ProRL-Agent-Server. Agentic RL on Any Harness at Scale

harbor. Framework for evaluating and improving agents

orbit. Stable and Efficient Reinforcement Learning for Trillion-Parameter LLMs

flashlib. Fast and memory-efficient classical machine learning operators

mKernel. mKernel: fast multi-node, multi-GPU fused kernels

substrate. Agent Substrate: the core system

motus. The open-source agent-serving project

HumanNet. HumanNet: Scaling Human-centric Video Learning to One Million Hours

lance. Open Lakehouse Format for Multimodal AI. Convert from Parquet in 2 lines of code for 100x faster random access, vector index, and data versioning. Compatible with Pandas, DuckDB, Polars, Pyarrow, and PyTorch with more integrations coming..

tokenspeed. TokenSpeed is a speed-of-light LLM inference engine.

pith-train. Compact and Agent-Native MoE Training System

humanize. From Automated Idea Factory to Realization

pretext. Fast, accurate & comprehensive text measurement & layout

uccl. UCCL is an efficient communication library for GPUs, covering collectives, P2P (e.g., KV cache transfer, RL weight transfer), and EP (e.g., GPU-driven)

hive. Python

autoresearch. AI agents running research on single-GPU nanochat training automatically

Reinforce-Ada. [COLM 2026] An adaptive sampling framework for Reinforce-style LLM post training.

VIGA. VIGA: Vision-as-Inverse-Graphics Agent

x-algorithm. Algorithm powering the For You feed on X

dflash. DFlash: Block Diffusion for Flash Speculative Decoding

mini-sglang. A compact implementation of SGLang, designed to demystify the complexities of modern LLM serving systems.

next-ai-draw-io. A next.js web application that integrates AI capabilities with draw.io diagrams. This app allows you to create, modify, and enhance diagrams through natural language commands and AI-assisted visualization.

flash-colreduce. Fast, memory-efficient attention column reduction (e.g., sum, mean, max)

cutile-python. cuTile is a programming model for writing parallel kernels for NVIDIA GPUs

SparseSpec. Accelerating Large-Scale Reasoning Model Inference with Sparse Self-Speculative Decoding

fouroversix. Code for the papers: “Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling” and “Adaptive Block-Scaled Data Types”

vortex_torch. Vortex: Programmable Sparse Attention for Agents as Algorithm Designers

miles. Miles is an enterprise-facing reinforcement learning framework for LLM and VLM post-training, forked from and co-evolving with slime.

flash-moba. C++

flashtensors. Python

kimi-cli. Kimi Code CLI is your next CLI agent.

ms-swift. Use PEFT or Full-parameter to CPT/SFT/DPO/GRPO 600+ LLMs (Qwen3.6, DeepSeek-V4, GLM-5.1, InternLM3, Llama4, ...) and 300+ MLLMs (Qwen3-VL, Qwen3-Omni, InternVL3.5, Ovis2.5, GLM4.5v, Gemma4, Llava, Phi4, ...) (AAAI 2025).

mini-swe-agent. The 100 line AI agent that solves GitHub issues or helps you in your command line. Radically simple, no huge configs, no giant monorepo—but scores >74% on SWE-bench verified!

inspect_ai. Inspect: A framework for large language model evaluations

skills. Public repository for Agent Skills

rlm-minimal. Super basic implementation (gist-like) of RLMs with REPL environments.

AHN. AHN: Artificial Hippocampus Networks for Efficient Long-Context Modeling

the-markovian-thinker. Code for paper "The Markovian Thinker: Architecture-Agnostic Linear Scaling of Reasoning"

tinker-cookbook. Post-training with Tinker

flash-kmeans. Fast and memory-efficient exact kmeans

cwm. Research code artifacts for Code World Model (CWM) including inference tools, reproducibility, and documentation.

TritonForge. 🔥 LLM-powered GPU kernel synthesis: Train models to convert PyTorch ops into optimized Triton kernels via SFT+RL. Multi-turn compilation feedback, cross-platform NVIDIA/AMD, Kernelbook + KernelBench

DeepResearch. Tongyi Deep Research, the Leading Open-source Deep Research Agent

batch_invariant_ops. Python

Parallel-R1. The offical repo for "Parallel-R1: Towards Parallel Thinking via Reinforcement Learning"

360-LLaMA-Factory. adds Sequence Parallelism into LLaMA-Factory

flex-nano-vllm. FlexAttention based, minimal vllm-style inference engine for fast Gemma 2 inference.

IMO25. An AI agent system for solving International Mathematical Olympiad (IMO) problems using Google's Gemini, OpenAI, and XAI APIs.

linear_cross_entropy_loss. A fusion of a linear layer and a cross entropy loss, written for pytorch in triton.

debugpy. An implementation of the Debug Adapter Protocol for Python

POLARIS. Scaling RL on advanced reasoning models

helion. A Python-embedded DSL that makes it easy to write fast, scalable ML kernels with minimal boilerplate.

Kimi-K2. Kimi K2 is the large language model series developed by Moonshot AI team

gemini-bball. Python

tritonparse. TritonParse: A Compiler Tracer, Visualizer, and Reproducer for Triton Kernels

sparselora. [ICML 2025] SparseLoRA: Accelerating LLM Fine-Tuning with Contextual Sparsity

genai-bench. Genai-bench is a powerful benchmark tool designed for comprehensive token-level performance evaluation of large language model (LLM) serving systems.

RL2. Python

verl-tool. A version of verl to support diverse tool use [TMLR 2026]

Compiler-R1. Compiler-R1: Towards Agentic Compiler Auto-tuning with Reinforcement Learning

torchtitan. A PyTorch native platform for training generative AI models

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

openevolve. Open-source implementation of AlphaEvolve

SandboxFusion. Python

CURE. [NeurIPS 2025 Spotlight] Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning

learn-cuda. Learn CUDA with PyTorch

memvid. Memory layer for AI Agents. Replace complex RAG pipelines with a serverless, single-file memory layer. Give your agents instant retrieval and long-term memory.

ROLL. An Efficient and User-Friendly Scaling Library for Reinforcement Learning with Large Language Models

good-kernels. Samples of good AI generated CUDA kernels

code-r1. Reproducing R1 for Code with Reliable Rewards

rllm. Democratizing Reinforcement Learning for LLMs

ParScale. Parallel Scaling Law for Language Model — Beyond Parameter and Inference Time Scaling

Absolute-Zero-Reasoner. Official Repository of Absolute Zero Reasoner

chat-ui. The open source codebase powering HuggingChat

open-webui. User-friendly AI Interface (Supports Ollama, OpenAI API, ...)

SkyRL. SkyRL: A Modular Full-stack RL Library for LLMs

grpo-optuna. Optimizing Causal LMs through GRPO with weighted reward functions and automated hyperparameter tuning using Optuna

ArcticInference. ArcticInference: vLLM plugin for high-throughput, low-latency inference

synthetic-data-kit. Tool for generating high quality Synthetic datasets

APR. [COLM 2025] Code for Paper: Learning Adaptive Parallel Reasoning with Language Models

FramePack. Lets make video diffusion practical!

KernelBench. KernelBench: Can LLMs Write GPU Kernels? - Benchmark + Toolkit with Torch -> CUDA (+ more DSLs)

DeepResearcher. Scaling Deep Research via Reinforcement Learning in Real-world Environments.

Triton-distributed. Distributed Compiler based on Triton for Parallel Systems

frontier-evals. OpenAI Frontier Evals

AReaL. The RL Bridge for LLM-based Agent Applications. Made Simple & Flexible.

dynamo. A Datacenter Scale Distributed Inference Serving Framework

atom. [NeurIPS 2025] Atom of Thoughts for Markov LLM Test-Time Scaling

Sparse-VideoGen. [ICML2025, NeurIPS2025 Spotlight] Sparse VideoGen 1 & 2: Accelerating Video Diffusion Transformers with Sparse Attention

R1-Searcher. R1-searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning

tilelang. Domain-specific language designed to streamline the development of high-performance GPU/CPU/Accelerators kernels

Search-R1. Search-R1: An Efficient, Scalable RL Training Framework for Reasoning & Search Engine Calling interleaved LLM based on veRL

DeepGEMM. DeepGEMM: clean and efficient BLAS kernel library on GPU

FlashMLA. FlashMLA: Efficient Multi-head Latent Attention Kernels

open-infra-index. Production-tested AI infrastructure tools for efficient AGI development and community-driven innovation

Open-Reasoner-Zero. Official Repo for Open-Reasoner-Zero

SWELancer-Benchmark. This repo contains the dataset and code for the paper "SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering?"

MoBA. MoBA: Mixture of Block Attention for Long-Context LLMs

simple_GRPO. A very simple GRPO implement for reproducing r1-like LLM thinking.

ReasonFlux. [NeurIPS 2025 Spotlight] LLM post-training suite — featuring ReasonFlux, ReasonFlux-PRM, and ReasonFlux-Coder.

RAGEN. RAGEN leverages reinforcement learning to train LLM reasoning agents in interactive, stochastic environments.

s1. s1: Simple test-time scaling

open-r1. Fully open reproduction of DeepSeek-R1

TinyZero. Minimal reproduction of DeepSeek R1-Zero

ColBERT. ColBERT: state-of-the-art neural search (SIGIR'20, TACL'21, NeurIPS'21, NAACL'22, CIKM'22, ACL'23, EMNLP'23)

maxtext. A simple, performant, and scalable Jax LLM!

verl. verl/HybridFlow: A Flexible and Efficient RL Post-Training Framework

mlx-examples. Examples in the MLX framework

nextjs-vllm-ui. Fully-featured, beautiful web interface for vLLM - built with NextJS.

bm25s. Fast BM25 search in Python, powered by Numpy and Numba

jute. A native Jupyter notebook frontend with local + remote kernels, reactive cells, and IDE features, implemented in Rust

Slow-Perception. Official code implementation of Slow Perception:Let's Perceive Geometric Figures Step-by-step

search-and-learn. Recipes to scale inference-time compute of open models

Slow_Thinking_with_LLMs. A series of technical report on Slow Thinking with LLM

SWE-Gym. Code for Paper: Training Software Engineering Agents and Verifiers with SWE-Gym [ICML 2025]

torchtune. PyTorch native post-training library

HunyuanVideo. HunyuanVideo: A Systematic Framework For Large Video Generation Model

open-instruct. AllenAI's post-training codebase

how-to-optim-algorithm-in-cuda. how to optimize some algorithm in cuda.

tiny-flash-attention. flash attention tutorial written in python, triton, cuda, cutlass

nano-sparse-attention. The simplest implementation of recent Sparse Attention patterns for efficient LLM inference.

SparTA. Python

rlt. Official Implementation for our NeurIPS 2024 paper, "Don't Look Twice: Run-Length Tokenization for Faster Video Transformers".

resource-stream. GPU programming related news and material links

Cosmos-Tokenizer. A suite of image and video neural tokenizers

nunchaku. [ICLR2025 Spotlight] SVDQuant: Absorbing Outliers by Low-Rank Components for 4-Bit Diffusion Models

COAT. [ICLR 2025] COAT: Compressing Optimizer States and Activation for Memory-Efficient FP8 Training

Video-XL. 🔥🔥First-ever hour scale video understanding models

depyf. depyf is a tool to help you understand and adapt to PyTorch compiler torch.compile.

tau-bench. Code and Data for Tau-Bench

llm-reasoners. A library for advanced large language model reasoning

finetrainers. Scalable and memory-optimized training of diffusion models

entropix. Entropy Based Sampling and Parallel CoT Decoding

awesome-o1. A bibliography and survey of the papers surrounding o1

hart. HART: Efficient Visual Generation with Hybrid Autoregressive Transformer

duo-attention. [ICLR 2025] DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads

FIFO-Diffusion_public. Official implementation of FIFO-Diffusion: Generating Infinite Videos from Text without Training (NeurIPS 2024)

EasyContext. Memory optimization and training recipes to extrapolate language models' context length to 1 million tokens, with minimal hardware.

Block-Sparse-Attention. A sparse attention kernel supporting mix sparse patterns

MiniCPM. MiniCPM5-1B: A SOTA 1B on-device LLM, small yet powerful.

flex_head_fa. Fast and memory-efficient exact attention

mle-bench. MLE-bench is a benchmark for measuring how well AI agents perform at machine learning engineering

LongRAG. Official repo for "LongRAG: Enhancing Retrieval-Augmented Generation with Long-context LLMs".

mirage. Mirage Persistent Kernel: Compiling LLMs into a MegaKernel

VideoSys. VideoSys: An easy and efficient system for video generation

fast-DiT. Fast Diffusion Models with Transformers

vllm-kvcompress. KV cache compression for high-throughput LLM inference

HELMET. The HELMET Benchmark

SageAttention. [ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.

FastV. [ECCV 2024 Oral] Code for paper: An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models

MathBlackBox. Python

xDiT. xDiT: A Scalable Inference Engine for Diffusion Transformers (DiTs) with Massive Parallelism

ReST-MCTS. ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search (NeurIPS 2024)

claude-cookbooks. A collection of notebooks/recipes showcasing some fun and effective ways of using Claude.

RAGatouille. Easily use and train state of the art late-interaction retrieval methods (ColBERT) in any RAG pipeline. Designed for modularity and ease-of-use, backed by research.

TESTA. [EMNLP 2023] TESTA: Temporal-Spatial Token Aggregation for Long-form Video-Language Understanding

Awesome-LLM-Strawberry. A collection of LLM papers, blogs, and projects, with a focus on OpenAI o1 🍓 and reasoning techniques.

g1. g1: Using Llama-3.1 70b on Groq to create o1-like reasoning chains

Vchitect-2.0. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models

LoRA-GA. Jupyter Notebook

LongWriter. [ICLR 2025] LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs

ell. A language model programming library.

lmms-eval. One-for-All Multimodal Evaluation Toolkit Across Text, Image, Video, and Audio Tasks

TiledCUDA. We invite you to visit and follow our new repository at https://github.com/microsoft/TileFusion. TiledCUDA is a highly efficient kernel template library designed to elevate CUDA C’s level of abstraction for processing tiles.

Qwen3-VL. Qwen3-VL is the multimodal large language model series developed by Qwen team, Alibaba Cloud.

hash-hop. Long context evaluation for large language models

Liger-Kernel. Efficient Triton Kernels for LLM Training

Visualizer. assistant tools for attention visualization in deep learning

mPLUG-Owl. mPLUG-Owl: The Powerful Multi-modal Large Language Model Family

TVTS. Turning to Video for Transcript Sorting

TempCompass. [ACL 2024 Findings] "TempCompass: Do Video LLMs Really Understand Videos?", Yuanxin Liu, Shicheng Li, Yi Liu, Yuxiang Wang, Shuhuai Ren, Lei Li, Sishuo Chen, Xu Sun, Lu Hou

MiniCPM-V. A Pocket-Sized MLLM for Ultra-Efficient Image and Video Understanding on Your Phone

clapper. Clapper.app, a video synthesizer and sequencer designed for the age of AI cinema

CogVideo. text and image to video generation: CogVideoX (2024) and CogVideo (ICLR 2023)

LLaVA-NeXT. Python

attention-gym. Helpful tools and examples for working with flex-attention

ring-flash-attention. Ring attention implementation with flash attention

cold-compress. Cold Compress is a hackable, lightweight, and open-source toolkit for creating and benchmarking cache compression methods built on top of GPT-Fast, a simple, PyTorch-native generation codebase.

Kangaroo. official impelmentation of Kangaroo: A Powerful Video-Language Model Supporting Long-context Video Input

MInference. [NeurIPS'24 Spotlight, ICLR'25, ICML'25] To speed up Long-context LLMs' inference, approximate and dynamic sparse calculate the attention, which reduces inference latency by up to 10x for pre-filling on an A100 while maintaining accuracy.

ktransformers. A Flexible Framework for Experiencing Heterogeneous LLM Inference/Fine-tune Optimizations

VLMEvalKit. Open-source evaluation toolkit of large multi-modality models (LMMs), support 220+ LMMs, 80+ benchmarks

CEPE. [ACL 2024] Long-Context Language Modeling with Parallel Encodings

temporal-shift-module. [ICCV 2019] TSM: Temporal Shift Module for Efficient Video Understanding

mem0. Universal memory layer for AI Agents

ao. PyTorch native quantization and sparsity for training and inference

SEA-RAFT. [ECCV2024 - Oral, Best Paper Award Candidate] SEA-RAFT: Simple, Efficient, Accurate RAFT for Optical Flow

distrifuser. [CVPR 2024 Highlight] DistriFusion: Distributed Parallel Inference for High-Resolution Diffusion Models

pie. Python

gan-compression. [CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs

CompressAI. A PyTorch library and evaluation platform for end-to-end compression research

kmcuda. Large scale K-means and K-nn implementation on NVIDIA GPU / CUDA

VideoTree. Code for CVPR25 paper "VideoTree: Adaptive Tree-based Video Representation for LLM Reasoning on Long Videos"

kmeans_pytorch. kmeans using PyTorch

ttt-lm-pytorch. Official PyTorch implementation of Learning to (Learn at Test Time): RNNs with Expressive Hidden States