Projects

### Overview TokenSwift is a framework designed to accelerate the generation of ultra-long sequences, specifically up to 100K tokens, while maintaining the quality of the original model. It achieves this by leveraging a novel approach that reduces computation time significantly. ### Key Features - **Speed**: 3× faster than vanilla Transformers. - **Lossless**: Matches the original model's output quality. - **Scalability**: Linear time complexity for 100K+ sequences. - **Plug & Play**: Works with most HuggingFace models. ### Tech Stack - Python ### Why It Matters TokenSwift is essential for applications requiring ultra-long sequence generation, such as language modeling, text completion, and more, by significantly reducing the time and computational resources required for these tasks. ### Repository - **Repository**: bigai-nlco/TokenSwift - **GitHub Stars**: 126 - **GitHub Description**: [ICML 2025] |TokenSwift: Lossless Acceleration of Ultra Long Sequence Generation ### 📖 Table of Contents - **Introduction** - **Installation** - **Method 1: With pip** - **Method 2: From the source (recommended)** - **Getting Started** - **Models Download** - **Inference** - **Training Guide (Optional)** - **Datasets Download** - **How to Train** - **Citation** - **Acknowledgment** ### Demo - **Demo Link**: https://github.com/user-attachments/assets/

Overview: This Chrome/Edge extension allows users to recognize and translate live audio and video streaming content for free, using an unofficial online Google Translate API. Key features: - Recognizes live audio and video streams - Translates content for free - Customizable fullscreen mode - Automatically applies settings changes in real time - Supports Stereo Mix or Virtual Audio Cable recording devices for better recognition Tech stack: JavaScript Why it matters: This extension provides a convenient and free way for users to translate live content, making it useful for language learners and those who need to quickly understand foreign language broadcasts. GitHub stars: 126 Repository: botbahlul/crx-live-translate

### Overview: 使用真实文化物证、精炼的中文标题和留白，东方文化编辑海报技能旨在创建简洁有力的视觉效果，适用于各种文化主题的海报、展览封面和出版物封面。 ### Key features: - **真实文化物证**：使用实物细节作为主题，增强视觉的真实感。 - **精炼标题**：通过标题动作和留白，使标题与图像产生关联，提升阅读体验。 - **A/B/C三联测试**：支持三张成组比较，方便快速验证和调整。 ### Tech stack: 使用图像生成作为生产步骤，支持直接图像海报和图像生成式海报版面。 ### Why it matters: 通过真实文化物证和精炼标题，东方文化编辑海报技能为设计者提供了一种简洁而富有文化内涵的视觉表达方式，适用于多种文化主题的海报设计，增强视觉效果和文化认同感。

### Overview: Explore the Limits of Omni-modal Pretraining at Scale ### Key features: - Collects large-scale omni-modal paired data including text, image, video, depth, and normal maps. - Designs an omni-modal learning architecture with two distinct branches for knowledge and interface modality. - Aligns the knowledge and interface modalities through a novel generative reasoning method. - Achieves general-purpose omni-modal understanding and learns transferable, universal representations. ### Tech stack: Python ### Why it matters: Inspired by cognitive neuroscience, this project aims to understand how the human brain processes multimedia information and how to leverage this knowledge to improve machine learning models. By collecting and pretraining omni-modal data, MiCo aims to achieve general-purpose understanding and transferable representations across different modalities, potentially revolutionizing the field of multimodal learning.

### Overview: InteractDiffusion is a pluggable interaction control model designed to extend existing pre-trained T2I diffusion models, enabling them to better respond to interactions in generated content. This project aims to enhance the ability of text-to-image diffusion models to handle interactions between objects, making them more versatile and effective in various applications. ### Key Features: - **Model Versioning:** Supports multiple versions (Tiny, Large, XLv1.0) with different training strategies. - **Interaction Control:** Enables conditional generation based on interactions, improving the realism and context of the generated images. - **Model Performance:** Measured by FID (Frechet Inception Distance) and KID (Koalas Inception Distance), with significant improvements in FID from v1.0 to v1.2. - **Web Interface:** Provides a web-based interface for users to interact with the model, enhancing usability and accessibility. ### Tech Stack: - Python ### Why It Matters: InteractDiffusion addresses a significant challenge in text-to-image diffusion models by providing a way to control interactions, making the generated images more realistic and contextually relevant. This model is particularly useful in applications where interactions between objects are important, such as in virtual environments, art, and creative design. ### Repository: - **Repository:** jiuntian/interactdiffusion - **GitHub Stars:** 126 - **GitHub Description:** Official repo for "InteractDiffusion: Interaction-Control for Text-to-Image Diffusion Model" at CVPR 2024. ### Primary

Overview: This project is an AI-powered task-driven autonomous agent system that utilizes GPT-4, Pinecone vector search, and the LangChain framework to efficiently complete, generate, and prioritize tasks. Key features: - Task Manager Agent - Task Executor Agent - Memory Manager Agent - Execution Context Agent - Security and Safety Agent Tech stack: Python, LangChain, GPT-4, Pinecone vector search. Why it matters: This system aims to enhance the efficiency and effectiveness of task management by leveraging advanced AI technologies, making it a valuable tool for developers and project managers. Repository: kalaspuff/ai-assisted-task-executor GitHub stars: 126 GitHub description: 🤖 A Task-Driven Autonomous Agent System as an AI-powered solution that leverages LLMs, vector search, and the LangChain framework to efficiently complete, generate, and prioritize tasks using a wide range of different concepts.

Overview: This project is a FastAPI boilerplate designed to kickstart your GenAI project, featuring a modular architecture, integrated tracing for debugging, production-ready deployment, and optimized performance for AI workflow orchestration. Key features: - 🧩 Modular LangGraph workflows - 🔍 Integrated Langfuse tracing - 🐳 Production-ready Docker deployment - 🧠 Optimized for AI workflow orchestration Tech stack: Python Why it matters: This boilerplate provides a clean, modular foundation for developing FastAPI applications, ensuring they are ready for production and scalable for AI-driven projects.

Overview: This project is a ChatGPT plugin for theorem proving in the Lean theorem prover, built using LeanDojo for backend development. It enables ChatGPT to interact with Lean, proving theorems by leveraging the power of language models and Lean's theorem proving capabilities. Key features: - Integrates ChatGPT with LeanDojo for theorem proving - Enables users to ask ChatGPT to prove theorems by providing the theorem name and its definition file - Provides a high-level proof plan and allows for various tactics to be used Tech stack: Python Why it matters: This project showcases the potential of integrating language models with theorem proving systems, providing a new way to enhance theorem proving capabilities.

Overview: Talk2BEV is a large vision-language model designed to enhance bird's-eye view (BEV) maps for autonomous driving tasks. It eliminates the need for BEV-specific training by leveraging pre-trained LVLMs, enabling a single system to perform various autonomous driving tasks such as visual and spatial reasoning, traffic actor intent prediction, and decision-making based on visual cues. Key features: - Eliminates the need for BEV-specific training - Enables a single system to cater to a variety of autonomous driving tasks - Supports visual and spatial reasoning - Predicts traffic actor intents - Decision-making based on visual cues Tech stack: The project uses Python as the primary language. Why it matters: Talk2BEV simplifies the development of autonomous driving systems by reducing the need for specialized BEV training data and enabling a single model to perform multiple tasks, making it more versatile and efficient. The release of Talk2BEV-Bench, a benchmark with 1000 human-annotated BEV scenarios, provides a valuable resource for further research in LVLMs for autonomous driving. Repository: The project is hosted on GitHub at https://github.com/llmbev/talk2bev. It has received 126 GitHub stars and is part of the ICRA 2024 conference.

Overview: Matryoshka Multimodal Models (M3) is a novel approach to learning visual tokens in a nested manner, following a coarse-to-fine order, for improved multimodal representation learning. Key features: - M3 proposes a hierarchical learning of visual tokens. - It is designed to be compatible with existing LLaVA models. - The model is trained on a variety of datasets and can be fine-tuned for specific tasks. Tech stack: Python Why it matters: M3 offers a more structured and efficient way to learn multimodal representations, potentially improving the performance of downstream tasks by leveraging the hierarchical nature of the visual token learning process. Repository: mu-cai/matryoshka-mm Primary language: Python GitHub stars: 126 GitHub description: Matryoshka Multimodal Models README excerpt: Matryoshka Multimodal Models [Paper](https://arxiv.org/abs/2405.17430) | [Project Page](https://matryoshka-mm.github.io/) | [Demo](https://pages.cs.wisc.edu/~mucai/matryoshka-mm.html) | [Model Zoo](https://github.com/mu-cai/matryoshka-mm/blob/main/docs/MODEL_ZOO.md) <!--p align="center"> <a href="https://llava.hliu.cc/"><img src="images/llava_logo.png" width="50%"></a> <br> Generated by <a href="https://gligen

Overview: This repository is a comprehensive project for CUDA programming, covering various advanced techniques such as GEMM, Flash Attention, Tensor Cores, CUTLASS, quantization, KV cache, NCCL, and performance optimization. It includes detailed code examples and real-world performance benchmarks. Key features: - GEMM (General Matrix-Matrix Multiplication) - Flash Attention - Tensor Cores - CUTLASS (CUDA Library for Tensor Core) - Quantization - KV Cache - NCCL (NVIDIA Communication Library) - Performance analysis and profiling Tech stack: CUDA C++ (C++17), CMake, cuBLAS, cuFFT, Thrust, WMMA, CUTLASS, NCCL Why it matters: This project provides a practical, hands-on approach to learning CUDA programming, focusing on optimizing performance and understanding the underlying hardware. It covers essential techniques for modern GPU computing, making it valuable for developers aiming to improve their CUDA skills. Repository: psmarter/CUDA-Practice GitHub stars: 127 GitHub description: A repository for CUDA programming practice, including GEMM, Flash Attention, Tensor Cores, CUTLASS, quantization, KV cache, NCCL, and performance optimization. It also includes real-world benchmarks and detailed documentation. Overview: This repository is a comprehensive project for CUDA programming, covering various advanced techniques such as GEMM, Flash Attention, Tensor Cores, CUTLASS, quantization, KV cache, NCCL, and performance optimization. It includes detailed code examples and real-world

Overview: AgentSmithy is a suite of tools and templates designed to simplify the process of building and deploying custom AI agents on Google Cloud. Key features: - Multi-agent runtime infrastructure - Industry-specific use cases - Predefined use cases by industry - Sub-agents categorized by industry (e.g., HCLS, FSI, etc.) - Central agent registry for mapping available agents and prompts Tech stack: Python Why it matters: AgentSmithy streamlines the development and deployment of AI agents, making it easier for developers to create and manage custom AI solutions on Google Cloud.

Overview: MinerU Skill is an AI-native document parser designed for AI agents, converting PDFs, Office documents, and images into clean Markdown with zero dependencies and API keys. Key features: - Zero dependencies - Fast parallel batches - Clean Markdown output - Supports PDF, Office, images, formulas, and tables - Agent-friendly output, including `--stdout` and `--json` status - Auto-routes to the Standard API for enhanced features - Supports multiple modalities and batch processing Tech stack: Python Why it matters: This tool simplifies the process of converting documents into Markdown format, making it easy for AI agents to access and utilize the information without requiring any additional dependencies or API keys. Repository: Nebutra/MinerU-Skill GitHub stars: 125

### Overview: My Wiki 是一个本地优先的知识管理系统，专为AI代理设计，提供CLI、可视化知识宇宙、Viki问答和可分享的知识宇宙。它简化了知识的保存、提取和应用过程，使AI代理能够直接访问和使用知识。 ### Key features: - **CLI 与知识维护**: 使用CLI进行知识的保存、提取和应用。 - **可视化知识宇宙**: 通过网页前端展示知识的结构和内容。 - **Viki 问答**: 通过Viki实现知识的问答功能。 - **本地知识库**: 知识保存在本地文件夹中，无需依赖云数据库或付费API。 - **Markdown 支持**: 知识库中的文件和图片均以Markdown格式保存。 - **Agent Skill**: 一个轻量级的适配层，允许在其他工作区中调用My Wiki项目。 ### Tech stack: JavaScript ### Why it matters: My Wiki简化了知识的保存、提取和应用过程，使AI代理能够直接访问和使用知识，提高了工作效率和知识的可访问性。它通过本地化和可扩展性，为开发者和企业提供了高效、安全的知识管理系统。

Overview: Nexus is a noise-decoupled prediction pipeline designed for adaptive driving scene generation, ensuring both timely reaction and goal-directed control. Key features: - Realistic safety-critical driving scenarios generated through flexible scene state control. - Utilizes NeRF for generating realistic environments. - Noise-decoupled prediction pipeline for adaptive scene generation. Tech stack: Python Why it matters: Nexus provides a robust solution for generating realistic driving scenes while maintaining safety and control, making it valuable for autonomous driving research and development. Repository: OpenDriveLab/Nexus Primary language: Python GitHub stars: 125 GitHub description: [ICCV 2025] Nexus: Decoupled Diffusion Sparks Adaptive Scene Generation README excerpt: > [!IMPORTANT] > 🌟 Stay up to date at [opendrivelab.com](https://opendrivelab.com/#news)! <div id="top" align="center"> <img src="https://opendrivelab.github.io/Nexus/resources/NEXUS-text.gif" alt="Image description" width="70%"> # Decoupled Diffusion Sparks Adaptive Scene Generation <!-- **Revive driving scene understanding by delving into the embodiment philosophy**--> <a href="http://arxiv.org/abs/2504.10485"><img src="https://img.shields.io/badge/arXiv-Paper-<color>"></a> <a href="https://opendrivelab.com/Nexus"><img src="https://img.shields.io

### Overview: Law Knowledge is a legal knowledge search and Q&A application based on Vietnam's Legal Code and legal document database. It aims to simplify legal research and provide users with relevant information quickly. ### Key Features: - 📖 Document Organization and Linking: Organizes legal documents and links them for easy navigation. - 📦 Terminology Extraction: Extracts legal terminology for improved search accuracy. - 🔍 Quick Search for Related Content: Allows users to quickly find related legal documents and information. - 🤖 User Query Support: Provides direct answers to user queries based on the legal knowledge database. ### Tech Stack: - TypeScript: The primary programming language used for the application. ### Why It Matters: Law Knowledge simplifies the process of legal research, making it more accessible and efficient for users. By leveraging the Vietnam's Legal Code and legal document database, it offers a comprehensive and reliable source of information for legal professionals and the general public. Repository: `foxminchan/LawKnowledge` Primary Language: TypeScript GitHub Stars: 124 GitHub Description: A legal knowledge search and Q&A application based on Vietnam's Legal Code and legal document database ⚖️

Overview: A groundbreaking project that introduces MaskUnet, a novel method for enhancing diffusion models by selectively masking parameters, achieving remarkable improvements in generation quality without additional parameters. Key features: - Improves diffusion generation quality by selectively masking U-Net parameters. - Supports both training-based and training-free optimization methods. - Achieves strong performance on COCO and downstream tasks with negligible additional parameters. Tech stack: Python Why it matters: This innovative approach leverages time-wise effective parameters to enhance diffusion models, offering a significant breakthrough in generative AI.

Overview: llm_profiler is a Python tool designed to analyze the theoretical performance of language models, providing insights into parameters, flops, memory, and latency. Key features: - Supports Qwen2.5 and Qwen3 dense series models. - Enables tensor parallel inference mode. - Supports hardware configurations including A100, V100, and T4. - Identifies memory-bound and compute-bound layers. - Outputs parameter, flops, memory, and latency for each layer and the entire model. - Calculates inference performance, including pre-filling and decoding latency. - Allows setting of computational efficiency and memory read efficiency. - Provides theoretical analysis results in a formatted output. Tech stack: Python Why it matters: llm_profiler offers valuable insights into the performance of language models, helping developers optimize their models for better efficiency and resource utilization. Repository: harleyszhang/llm_counts Primary language: Python GitHub stars: 124

Overview: 小女仆机器人是一个轻量级、自托管的多入口AI代理机器人，支持QQ官方、OneBot 11和微信服务号的文本入口，提供聊天、记事、工具调用和主动推送等功能。它使用Rust语言构建，具备多轮会话、受控长期记忆、Todo、RSS推送、本地知识检索、联网查询、多模态理解以及Tool Calling等特性。 Key features: - 轻量级自托管 - 支持QQ官方、OneBot 11和微信服务号的文本入口 - 多轮会话 - 受控长期记忆 - Todo功能 - RSS推送 - 本地知识检索 - 联网查询 - 多模态理解 - Tool Calling - 主动推送 Tech stack: Rust Why it matters: 小女仆机器人是一款轻量级、自托管的多入口AI代理机器人，支持QQ官方、OneBot 11和微信服务号的文本入口，提供聊天、记事、工具调用和主动推送等功能。它使用Rust语言构建，具备多轮会话、受控长期记忆、Todo、RSS推送、本地知识检索、联网查询、多模态理解以及Tool Calling等特性。

Overview: OneRAG is a production-ready RAG framework built with Python and FastAPI. It supports multiple vector databases (Weaviate, Pinecone, Qdrant, ChromaDB, MongoDB), five language models (Gemini, OpenAI, Claude, Ollama, OpenRouter), and is compatible with the OpenAI API. Key features: - Configurable with 1-line changes for 6 vector databases - Configurable with 5 language models - Compatible with OpenAI API - 2100+ tests Tech stack: Python/FastAPI Why it matters: OneRAG simplifies the integration of various RAG components, making it easy to switch between different providers, models, and databases. It supports OpenAI-compatible APIs, ensuring compatibility with existing OpenAI code, and provides comprehensive testing coverage. Repository: notadev-iamaura/OneRAG Primary language: Python GitHub stars: 125 GitHub description: Production-ready RAG Framework (Python/FastAPI). 1-line config swaps: 6 Vector DBs (Weaviate, Pinecone, Qdrant, ChromaDB, pgvector, MongoDB), 5 LLMs (Gemini, OpenAI, Claude, Ollama, OpenRouter). OpenAI-compatible API. 2100+ tests. README excerpt: <p align="center"> <img src="assets/logo.svg" alt="OneRAG Logo" width="400"/> </p> <p align="center"> <strong>