Projects

Overview: TurboDiffusion is a video generation framework that accelerates diffusion models by up to 200 times, making it 100 times faster on a single RTX 5090 GPU while maintaining high video quality. Key features: - TurboDiffusion provides a 100-200x acceleration for video diffusion models. - Uses SageAttention, SLA, and rCM for attention acceleration and timestep distillation. - Supports both Wan2.1 and Wan2.2 models with checkpoints available. - Available in HuggingFace models and source code. Tech stack: Python, PyTorch, SageAttention, SLA, rCM. Why it matters: TurboDiffusion significantly speeds up the generation of videos, making it more practical for real-time applications and faster iterations in research.

### Overview: This repository contains a dataset of 15,140 jailbreak prompts collected from various platforms, including Reddit, Discord, websites, and open-source datasets. The dataset includes 1,405 jailbreak prompts, making it the largest collection of in-the-wild jailbreak prompts. ### Key Features: - 15,140 jailbreak prompts collected from December 2022 to December 2023. - 1,405 jailbreak prompts identified from various platforms. - Data source includes Reddit, Discord, websites, and open-source datasets. - Prompts are collected from four platforms: Reddit, Discord, websites, and open-source datasets. ### Tech Stack: JailbreakHub ### Why It Matters: This dataset serves as the largest collection of in-the-wild jailbreak prompts, providing valuable insights into the characteristics and effectiveness of jailbreak prompts in large language models.

### Overview: 本项目收集了在寒假期间学习《动手学习深度学习》过程中详细的Markdown笔记和相关的Jupyter代码，旨在帮助大家一边听课一边理解，同时提供详细的中文注释和代码示例，以提高实践能力。 ### Key features: - Markdown笔记与原课程视频一一对应，帮助理解课程内容。 - Jupyter代码有详细中文注释，便于快速上手实践。 - 包含深度学习模型的代码实现，包括LeNet、ResNet、LSTM和BERT等。 - 教材目录和章节链接，提供教材资源。 ### Tech stack: Jupyter Notebook ### Why it matters: - 提供了深度学习基础模型和前沿知识的实践机会。 - 便于理解和学习深度学习，特别是对于机器学习基础较弱的读者来说非常有用。 - 通过代码实现，可以快速上手并掌握深度学习的实际应用。

Overview: This is a Python port of the Fréchet Inception Distance (FID) measure to PyTorch, used to evaluate the quality of samples from Generative Adversarial Networks (GANs). It calculates the Fréchet distance between two Gaussians fitted to feature representations of the Inception network. Key features: - Computes FID scores with PyTorch - Supports evaluation on GPU - Provides compatibility with different feature layers of the Inception network - Generates a `.npz` archive for saving multiple model training times Tech stack: Python Why it matters: The FID score is a useful metric for comparing the quality of images generated by GANs, and this implementation provides a convenient way to compute it using PyTorch, making it accessible for developers and researchers alike. Repository: mseitzer/pytorch-fid GitHub stars: 3850

Overview: MatchZoo is a comprehensive toolkit designed for the design, comparison, and sharing of deep text matching models. Key features: - Unified data processing pipeline - Simplified model configuration - Automatic hyper-parameter tuning Tech stack: Python Why it matters: MatchZoo simplifies the process of designing, comparing, and sharing advanced deep learning models for text matching tasks such as document retrieval, question answering, conversational response ranking, and paraphrase identification. It provides a flexible and user-friendly environment for researchers and practitioners alike. Repository: NTMC-Community/MatchZoo Primary language: Python GitHub stars: 3848 GitHub description: Facilitating the design, comparison and sharing of deep text matching models.

Overview: TensorFlowOnSpark brings TensorFlow programs to Apache Spark clusters, enabling distributed deep learning on GPU and CPU servers. Key features: - TensorFlowOnSpark supports TensorFlow functionalities including synchronous/asynchronous training, model/data parallelism, and inferencing. - It leverages TensorFlow's built-in APIs for reading data files directly from HDFS and Spark RDDs for data ingestion. - The API allows seamless integration with existing Spark data processing pipelines. - TensorFlowOnSpark supports server-to-server direct communication for faster learning. Tech stack: Python Why it matters: TensorFlowOnSpark simplifies the migration of existing TensorFlow programs to a distributed deep learning environment on Apache Spark clusters, minimizing code changes and leveraging existing TensorFlow functionalities.

### Overview: Habitat-Sim is a high-performance physics-enabled 3D simulator designed for embodied AI research. It supports various datasets and configurations, including 3D scans, CAD models, and configurable sensors. The simulator achieves impressive performance, rendering scenes at up to 10,000 FPS on a single GPU and simulating tasks at high speeds, making it ideal for training and benchmarking agents in embodied AI experiments. ### Key features: - **3D Scans and Datasets Support**: Supports indoor and outdoor scenes from datasets like HM3D, HSSD, MatterPort3D, Gibson, and Replica. - **CAD Models and Piecewise-Rigid Objects**: Handles CAD models and piecewise-rigid objects like ReplicaCAD and YCB. - **Configurable Sensors**: Capable of rendering with RGB-D cameras and egomotion sensing. - **Robots via URDF**: Supports mobile manipulators, fixed-base arms, and quadrupeds like Fetch, Franka, and AlienGo. - **Rigid-Body Mechanics**: Utilizes Bullet physics for accurate simulation of rigid-body dynamics. ### Tech stack: C++ ### Why it matters: Habitat-Sim is crucial for researchers and developers in the field of embodied AI, providing a powerful tool for simulating complex environments and tasks. Its high performance and versatility make it an essential resource for advancing the capabilities of AI agents in real-world scenarios. ### Repository: facebookresearch/habitat-sim ### GitHub stars: 3846 ### GitHub description: A flexible, high

### Overview: Tutti is a multi-user, multi-agent, real-time collaboration space where people and agents can build together in harmony. ### Key features: - **Real-time Collaboration**: Agents can connect and collaborate in real-time. - **Built-in Apps**: A variety of apps are available for easy integration. - **Group Chat**: Facilitates communication among multiple users. - **Work with Others**: Agents can work together on projects. - **Work with Others' Agents**: Collaborate with agents from other users. - **Simultaneous Editing**: Multiple agents can edit the same document simultaneously. - **Agent Borrowing**: Agents can borrow and use others' agents. - **Across Devices**: Supports collaboration across different devices. - **No-Deploy Sharing**: Agents can share their work without deploying it. - **Big @**: Agents can collaborate across multiple users. - **Download**: Available for both open-source and VM versions. - **Creating a Room**: Requires an invite code for VM. ### Tech stack: TypeScript ### Why it matters: Tutti is a groundbreaking platform that enables seamless collaboration between people and agents, making it an essential tool for modern teamwork and real-time collaboration. ### Repository: tutti-os/tutti ### GitHub stars: 3846 ### GitHub description: Where people and agents build in tune.

Overview: Maestro is a highly scalable and extensible workflow orchestrator designed by Netflix to manage and schedule thousands of workflows and millions of jobs daily, ensuring strict service level objectives (SLOs) even during traffic spikes. Key features: - Highly scalable architecture - Extensible to support new and existing use cases - Enhanced usability for end users - Supports various workflows and job types Tech stack: Java Why it matters: Maestro is crucial for Netflix's data and machine learning workflows, enabling efficient data processing and management across the organization.

### Overview: Scenic is a JAX library designed for research in attention-based models for computer vision, including classification, segmentation, and detection across multiple modalities such as images, video, audio, and multimodal combinations. ### Key features: - **Light-weight Libraries**: Provides shared libraries for launching experiments, summary writing, logging, and profiling. - **Training and Evaluation Loops**: Optimized training and evaluation loops, losses, metrics, and bi-partite matchers. - **Input Pipelines**: Input pipelines for popular vision datasets. - **Baseline Models**: Baseline models such as ViViT, OmniNet, and TokenLearner. ### Tech stack: Developed in JAX and uses Flax. ### Why it matters: Scenic is essential for researchers and developers looking to accelerate research in computer vision by providing a unified framework for training and evaluating large-scale vision models.

### Overview: 零到NLP：中文NLP解决方案，基于PyTorch和Transformers，提供从数据处理到模型训练、推理的全套解决方案，适用于大规模数据和多模态任务。 ### Key features: - 🎯 目标：基于PyTorch和Transformers构建中文领域的开箱即用训练框架，涵盖数据处理、模型训练、微调、推理等全流程。 - 💽 数据：整理了海量训练数据，提供数据清洗和处理模板，支持多线程和内存映射处理大规模数据。 - 💻 流程：完整模型训练步骤，包括数据清洗、数据处理、模型构建、模型训练、模型部署、模型图解。 - 🔥 模型：支持多种大模型，包括GPT2、CLIP、Dolly、Llama、ChatGLM-6B等。 - 🚀 多卡串联：训练时和推理时多卡串联功能，支持部分模型结构修改以适应更大模型尺寸。 - ⚙️ 模型工具：提供了词表裁切和扩充教程。 ### Tech stack: Python, PyTorch, Transformers, Jupyter Notebook ### Why it matters: 零到NLP提供了一个全面的中文NLP解决方案，适用于各种任务，从数据处理到模型训练和推理，简化了开发过程，加速了项目启动和部署。

### Overview: SimpleMem is a text and multimodal memory solution designed for efficient lifelong learning in large language models (LLMs). It offers semantic lossless compression for storing and retrieving long-term memories, making it ideal for applications requiring continuous learning and adaptation. ### Key Features: - **Semantic Lossless Compression**: Efficiently stores and retrieves long-term memories with minimal data loss. - **Multimodal Support**: Integrates text, image, audio, and video for comprehensive memory storage and retrieval. - **Python Integration**: Seamless integration with Python environments, including support for MCP (text memory) and Python-based inference. ### Tech Stack: - Python ### Why It Matters: SimpleMem simplifies the process of lifelong memory management in LLMs, enabling continuous learning and adaptation without the need for complex vendor setups. Its efficient memory compression and multimodal support make it a valuable tool for developers and researchers working with AI models.

Overview: PipesHub is an open-source platform that connects AI agents to your company's knowledge systems, such as Slack, Google Drive, GitHub, and Microsoft 365, enabling them to search, grep, navigate, and cite information across multiple platforms. Key features: - Connects to over 50+ systems including Slack, Google Drive, GitHub, and Microsoft 365. - Provides permission-aware workspace for agents to search, grep, navigate, and cite information. - Built-in agents and SDKs for easy integration. - Self-hosted solution for secure, reliable, and scalable deployment. Tech stack: Python Why it matters: PipesHub empowers AI agents to access and analyze information from various business systems, making it easier to extract insights and improve decision-making processes.

Overview: 本项目提供了一套深入浅出的PyTorch入门教程，旨在帮助开发者掌握PyTorch工具，通过组队学习实现从入门到熟练的全过程。 Key features: - 第零章：前置知识（选学） - 第一章：PyTorch的简介和安装 - 第二章：PyTorch基础知识 - 第三章：PyTorch的主要组成模块 - 第四章：PyTorch基础实战 - 第五章：PyTorch模型定义 - 第六章：PyTorch进阶训练技巧 - 第七章：PyTorch可视化 - 第八章：PyTorch生态简介 - 第九章：模型部署 - 第十章：常见网络代码的解读 Tech stack: Jupyter Notebook Why it matters: 通过组队学习，开发者能够掌握由浅入深的PyTorch基本知识和内容，通过动手实践加深操作的熟练度，提升解决实际问题的能力。

### Overview: This repository contains the notes and notebooks created by Andrew Ng for his Coursera course on deep learning. The repository includes a collection of articles, notebooks, and videos that cover various aspects of deep learning, neural networks, and machine learning projects. ### Key Features: - **Articles:** - **[Neural Network Deep Learning](https://github.com/ashishpatel26/Andrew-NG-Notes/blob/master/andrewng-p-1-neural-network-deep-learning.md)** - **[Improving Deep Learning Network](https://github.com/ashishpatel26/Andrew-NG-Notes/blob/master/andrewng-p-2-improving-deep-learning-network.md)** - **[Structure of ML Projects](https://github.com/ashishpatel26/Andrew-NG-Notes/blob/master/andrewng-p-3-structuring-ml-projects.md)** - **[Convolutions Neural Network](https://github.com/ashishpatel26/Andrew-NG-Notes/blob/master/andrewng-p-4-convolutional-neural-network.md)** - **[Sequence Models](https://github.com/ashishpatel26/Andrew-NG-Notes/blob/master/andrewng-p-5-sequence-models.md)** - **Notebooks:** - **Week1 - Introduction to deep learning** - **Week2 - Neural Networks** - **Week3 - Deep Neural Networks**

### Overview: LoRAX is a multi-LoRA inference server designed to scale efficiently to serve thousands of fine-tuned Large Language Models (LLMs) on a single GPU, significantly reducing the cost of deployment without compromising performance. ### Key Features: - **Dynamic Adapter Loading:** Load any fine-tuned LoRA adapter from HuggingFace, Predibase, or local filesystems, enabling just-in-time loading without blocking requests. - **Heterogeneous Continuous Batching:** Pack requests for different adapters into the same batch, maintaining consistent latency and throughput. - **Adapter Exchange Scheduling:** Prefetch and offload adapters between GPU and CPU memory, optimizing request batching for system throughput. - **Optimized Inference:** Enhances performance with tensor parallelism, pre-compiled CUDA kernels, quantization, and token streaming. - **Ready for Production:** Prebuilt Docker images, Helm charts for Kubernetes, Prometheus metrics, and distributed tracing with OpenTelemetry. Supports OpenAI API for multi-turn chat conversations. Private adapters via per-request tenant isolation. - **Free for Commercial Use:** Apache 2.0 License, making it suitable for both research and commercial applications. ### Tech Stack: Python ### Why It Matters: LoRAX simplifies the deployment of multiple fine-tuned LLMs by efficiently managing and scaling them, making it ideal for applications requiring high throughput and low latency, such as chatbots, language models, and more.

### Overview: This repository is a comprehensive collection of public datasets and related studies in the field of industrial image anomaly/defect detection. It aims to provide a detailed overview and insights into the latest advancements in this area, encouraging readers to explore and contribute to the field. ### Key features: - **Datasets and Studies**: A curated list of public datasets and related research papers in industrial image anomaly/defect detection. - **Code Examples**: Links to open-source code repositories for various anomaly detection methods. - **Benchmarking**: A benchmarking tool for industrial image anomaly detection, including datasets and evaluation metrics. ### Tech stack: Python, Jupyter Notebook, and various machine learning libraries. ### Why it matters: This repository serves as a valuable resource for researchers, practitioners, and developers interested in industrial image anomaly detection. It provides a clear overview of the latest advancements, enabling easy comparison and learning from existing work. ### Repository: M-3LAB/awesome-industrial-anomaly-detection ### GitHub stars: 3832 ### GitHub description: Paper list and datasets for industrial image anomaly/defect detection (updating).

Overview: Implicit is a Python library designed for fast collaborative filtering on implicit feedback datasets, providing implementations of Alternating Least Squares, Bayesian Personalized Ranking, Logistic Matrix Factorization, and Item-Item Nearest Neighbour models. It includes multi-threaded training routines and CUDA kernels for GPU acceleration, making it suitable for applications requiring high performance. Key features: - Alternating Least Squares - Bayesian Personalized Ranking - Logistic Matrix Factorization - Item-Item Nearest Neighbour models - Multi-threaded training routines - Cython and OpenMP support - CUDA kernels for GPU acceleration - Approximate nearest neighbours libraries for recommendation speed up Tech stack: Python Why it matters: Implicit is essential for developers looking to build fast and efficient recommendation systems for implicit feedback datasets, making it a valuable tool for applications in music recommendation, fashion recommendations, and more. Repository: benfred/implicit GitHub stars: 3830

### Overview: 本项目将《动手学深度学习》(Dive into Deep Learning)原书中的MXNet实现改为TensorFlow 2.0实现，项目已得到李沐老师的认可。原书作者包括阿斯顿·张、李沐、扎卡里 C. 立顿、亚历山大 J. 斯莫拉以及其他社区贡献者，GitHub地址：https://github.com/d2l-ai/d2l-zh。此项目面向对深度学习感兴趣，尤其是使用TensorFlow 2.0进行深度学习的开发者，无需任何深度学习或机器学习背景知识。 ### Key features: - TensorFlow 2.0实现 - 改进的代码和文档 - 项目已得到李沐老师的认可 ### Tech stack: Jupyter Notebook ### Why it matters: 本项目为学习和实践深度学习提供了一个实用的资源，特别是对于使用TensorFlow 2.0进行开发的开发者来说，它简化了代码实现并提供了更好的文档体验。

Overview: DeepVariant is a deep learning-based variant caller designed for analyzing next-generation DNA sequencing data, specifically for calling germline variants in diploid organisms. It processes aligned reads, generates pileup image tensors, and classifies them using a convolutional neural network to produce a standard VCF or gVCF file. Key features: - Calls germline variants in diploid organisms - Supports both whole genome and whole exome data - Handles PacBio HiFi, Oxford Nanopore, Complete Genomics, and Roche SBX data - Provides case studies for various sequencing technologies and pangenome-aware mapping - Supports pangenome-aware variant calling for WGS and WES data Tech stack: Python Why it matters: DeepVariant leverages deep learning to accurately identify genetic variants from high-throughput sequencing data, making it a valuable tool for genetic research and clinical applications. Repository: google/deepvariant Primary language: Python GitHub stars: 3825 GitHub description: DeepVariant is an analysis pipeline that uses a deep neural network to call genetic variants from next-generation DNA sequencing data.