Schedule: September 14

Rust for Linux

From Go to Rust: Rust's Implementation in Tbps-level Live Streaming Protocol Stack

Industrial-grade Rust Verification

This presentation introduces Qingcheng.ai's design and development of a training and inference optimization platform for indigenous large models, with improving indigenous intelligent computing power as the core application scenario. The platform adopts the self-developed Chitu inference engine, designs a unified framework for training and fine-tuning, and proposes hybrid precision and low-bit quantization optimization schemes. While maintaining computational accuracy and general problem-solving capabilities, it significantly improves model computational speed and reduces computational resource consumption through engineering and product practices.

Cosmos World Foundation Models for Physical AI

Personal AI agents are one of the most exciting areas in AI. They have the potential to assist individuals in thousands of personal tasks, projects, and decisions. But for personal AI agents to take action on behalf of individuals, they must have the explicit trust and permission of their principals. This presentation explores the fundamental challenges of establishing trust relationships between humans and AI agents, including issues of consent, delegation of authority, privacy protection, and accountability. We examine technical solutions for trust frameworks, cryptographic proof systems, and governance models that can enable safe and beneficial personal AI agent deployment.

This presentation focuses on the evolution journey of the kOS Agentic System, exploring the development directions and implementation paths of next-generation autonomous intelligent systems. We examine the architectural innovations that enable truly autonomous decision-making, the integration of multi-modal AI capabilities, and the challenges of creating systems that can operate independently while maintaining safety and reliability. The talk covers practical applications, system design principles, and the future roadmap for autonomous intelligent platforms.

The 2024 Nobel Prize in Physics was awarded to Hopfield and Hinton for their foundational discoveries and inventions that enable machine learning through artificial neural networks. The 2024 Nobel Prize in Chemistry was awarded to David Baker, Hassabis, and Jumper for their contributions to protein design and prediction. These major events mark the arrival of the human intelligence age. This presentation explores the paradigm shift from individual innovation to collaborative intelligence, examining how AI systems and human creativity can work together to address complex global challenges and drive unprecedented scientific breakthroughs.

SGLang is an efficient open-source framework for large-scale LLM serving. Over the past year, SGLang has experienced rapid iteration and significant advancements. This presentation presents an overview of SGLang's leading features, including KV Cache Reuse, Zero-overhead Batch Scheduling, Speculative Decoding, Prefill/Decode Disaggregation, and Large-scale Expert Parallelism.

This presentation introduces the intrinsic advantages of the dora-rs framework and its concrete applications across multiple embodied robotics platforms — exemplified by dora-rs-enabled grasping and sorting systems.

Opening Remarks

The core of open source is institutional innovation. Whether it's the United States, European Union, or China, the obstacles to open source are the traditional thoughts, cultures, values, and rule systems formed over thousands of years based on the exclusive properties of the physical world. Additionally, countries around the world urgently need secure and reliable digital public goods and services in developing their digital economies and digital transformation, which far exceeds the scope of existing international organizational systems established based on geographical spaces, requiring new technologies, new thinking, new models, and broader international cooperation.

Educational Ecological Changes from the Perspective of Educational Elements

Introduction to Dora's technical architecture, robotics cases developed based on Dora, and the construction of the Dora Chinese community.

In this age of limitless digital possibilities, open source is not only a symbol of technological innovation but also a key driver of global business expansion. This talk will take you through the journey of open source projects from inception to crossing borders and entering global markets. As a founder who started in the open source community and successfully turned it into a global business, I will share how open source innovation can break regional barriers and gain the support of global enterprises and investors. We will explore how to build a sustainable business model using open source, tackle challenges in different markets, and commercialize globally. Through real-life case studies, I will analyze how to turn technological passion and community support into long-term business success.

Single-source cross-platform GPU scientific computing with Slang and Rust

The Transformation of Teachers in the AI Era: From Knowledge Communicators to Knowledge Managers

vLLM is a fast and easy-to-use library for LLM inference and serving. In this talk, I will briefly introduce the evolution of the vLLM project, the open-source community behind it, and highlight some features that are interesting to many users.

Embodied Cerebrum-Cerebellum Operation Framework and Brain Model Construction

The Agentic Web needs digital identity that is independent of central authorities and intermediaries. This talk will present the technology, current activities, and the community of Decentralized Identifiers (DIDs), which enable verifiable, self-sovereign digital identity for AI agents and autonomous systems. We explore how DIDs can provide secure, privacy-preserving identity management for agent-to-agent interactions, the challenges of implementing decentralized identity in distributed AI systems, and the role of blockchain and cryptographic technologies in ensuring trust and authenticity in the agentic web ecosystem.

The Construction of AI Personalized Memory Operating System

Sovereign AI at Scale: Data Pipelines and Research Infrastructures for European Foundation Models

Large Language Model (LLM) inference comprises two distinct phases: Prefill and Decode. The Prefill phase is computation-intensive, processing the entire input sequence, while the Decode phase is memory-intensive, managing the Key-Value (KV) cache for token generation. Traditionally, these phases are handled within a unified engine, where combined scheduling of prefill and decode batches introduces inefficiencies. To address these challenges, we introduce Prefill and Decoding (PD) Disaggregation in SGLang, which enables tailored optimizations for each.

Adora Robot & Huawei Ascend-Based Dora Data Platform

Many Chinese open-source/AI projects have gained some recognition domestically but struggle to go global. This talk will reveal the four major misconceptions about open source globalization, analyze the main global battlefields and trends for 2025, and share how to make your project stand out among international developers through community operations and strategic planning, using real case studies.

Access to high-quality tutoring remains out of reach for millions of students due to cost, geography, and limited availability of skilled teachers. This talk presents a vision for an affordable, AI-powered STEM tutoring platform that can deliver personalized, adaptive learning experiences to students worldwide. We explore the technical architecture behind scalable AI tutoring systems, the challenges of creating culturally sensitive educational content, and the practical implementation strategies for making high-quality STEM education accessible to underserved communities globally.

Component-based Kernel Design Methods and Key Technology Research Based on Rust

Integration of Dora with Isaacsim

Open source makes AI development faster and more open, but it also brings responsibility challenges, such as model bias and compliance risks. In this sharing, I will talk about how the open source community collaborates and how we should find a balance between openness and responsibility.

Introduction to Ant Group's Ming-Omni series open-source work in the large model field, sharing practices and progress in multimodal large model architecture evolution, cross-modal fusion, and unified generation and understanding. Through joint design and optimization of model architecture and training processes, we are committed to building comprehensive modal capabilities for models, achieving multimodal foundation models that can see, hear, speak, and draw.

An overview of a $500 open-source humanoid arm and glove system enabling real-time learning and teleoperation. We'll explore how AI, robotics, and embedded systems merge to empower individual creators and researchers with accessible humanoid robotics technology. This presentation covers the design philosophy, technical architecture, manufacturing considerations, and community development model that makes advanced robotics accessible to makers, students, and researchers worldwide. We examine the integration challenges, control systems, and the potential for democratizing humanoid robotics development.

Exploring and Practicing Agentic WebApp Development

Second Me aims to train a personal large model shaped by your memories. It uses your views and experiences (as well as your daily recorded thoughts) as core training data, making AI no longer just a knowledge base, but something that can carry your expression, preferences, and boundaries. We will demonstrate the thinking behind Second Me, practical experiences during development, the use of Second Me in actual products, and some issues that need to be explored.

As large model parameters and context windows continue to expand simultaneously, dense attention computation has become the absolute bottleneck for online inference—memory usage follows a quadratic curve, and latency scales linearly with sequence length. The industry commonly adopts KVCache + speculative decoding to mitigate this, but it still falls short in extreme scenarios with ultra-long sequences. This sharing focuses on the "sparse attention" paradigm, systematically presenting for the first time our self-developed UCM inference memory data manager's sparse inference stack architectural design and implementation experience, from algorithms and plugin design to software implementation.

Technical introduction of iFLYTEK's PD (Prefill-Decode) separation approach based on open-source engines combined with their own inference service framework.

Roundtable meeting: Future School Paradigm: Human-Machine Symbiotic Educational Ecosystem

This session shares common challenges and practical recommendations for Dora development on OpenHarmony. Key topics include: development environment setup, cross-compilation strategies, hardware adaptation methods, debugging techniques, and performance optimization. The presentation will provide hands-on guidance for developers looking to implement Dora-based robotics solutions on the OpenHarmony platform, covering both technical implementation details and practical deployment considerations.

A roundtable discussion on the opportunities and challenges of open source globalization.

Building MCP Server and Client with Rust: Injecting Powerful Engines into Agents

Unleashing GenAI Inference at Scale: A Deep Dive into the llm-d Project

Modern robotics, V2V, and autonomous systems need fast, scalable, and resilient communication. Zenoh unifies pub/sub, queries, and storage, offering low latency, Tier-1 ROS 2 middleware rmw_zenoh, and seamless edge-to-cloud integration. This presentation explores how Zenoh revolutionizes robotics communication architectures, enabling distributed systems that can scale from single robots to massive fleets while maintaining real-time performance and fault tolerance. We examine practical deployment scenarios, performance benchmarks, and the roadmap for next-generation robotics infrastructure.

TMCP/TA2A: Running MCP and A2A over TSP

This talk explores the feasibility and specific implementation of Large Language Models (LLMs) as auxiliary combat agents in multi-unit Real-Time Strategy (RTS) games. Unlike traditional rule-based or heuristic AI systems, LLM-based agents can understand and parse players' high-level commands and abstract language through natural language understanding, and dynamically convert them into executable game operations. We analyze key issues in the implementation process, including player semantic understanding, command learning comprehension, game structure learning, game state understanding, execution content output, and error correction, as well as challenges in real-time performance and accuracy trade-offs.

Towards Biological General Intelligence

Enabling LLMs to improve their outputs with reinforcement learning and test-time scaling (TTS) is now a tendency, which leads to the adoption and unprecedented importance of inference engines in model training. In this presentation, we will go through the recent progress and community adoption in SGLang under RL scenarios.

DoRobot: A Containerized Embodied Intelligent Operating System Integrating Dora, ROS and LeRobot

Embedded Rust for AI

In this workshop, we will teach you how to build a simple app with Makepad, a modern GPU-based UI Framework for Rust that allows for writing fast and lightweight apps. The app we are going to build is a simple but functional demonstration of Makepad's capabilities, showcasing its unique approach to GPU-accelerated rendering and declarative UI design. Participants will learn the fundamentals of Makepad's architecture, how to create responsive layouts, handle user interactions, and leverage Rust's performance benefits for creating smooth, native-feeling applications across platforms.

Unified Index Architecture in Rust: Engineering Practice from Heterogeneous Indexes to Unified Abstraction

Robrix: a complex multi-platform Rust app for secure Matrix chat

LanceDB: The AI-Native Multimodal Lakehouse

Currently, SGLang exclusively supports the PyTorch AI framework and runs on Ascend NPUs through the torch_npu plugin. We designed a dual-framework collaborative solution for PyTorch and MindSpore, fully considering cross-framework resource reuse while simplifying interfaces to leverage MindSpore's performance advantages on Ascend NPUs. Developers can easily integrate MindSpore models into SGLang while maintaining unchanged user workflows. MindSpore's integration solution also provides a foundation for subsequent integration of other frameworks.

How to Migrate a Project from ROS to Dora

OAKS ASKG: Agent-Server Knowledge Graph

Application of Rust in the MoonBit Toolchain and Comparison of Both in Wasm Application Scenarios

How Doctests Work

GenUI: Rust Frontend Framework and AI Code Generation

Accessible Agentic Computer Vision with Kornia

Self-Sovereign Identity and Trust Infrastructure for Agentic IA in Data Spaces

This presentation focuses on the core challenges and key opportunities of spiking computing, an emerging brain-inspired computing paradigm. Spiking computing is characterized by event-driven and asynchronous information processing, distinguishing it from traditional continuous synchronous computing. The presentation first addresses core challenges: the difficulties in constructing discontinuous computing models, exploring how to establish efficient and practical algorithmic frameworks, followed by an exposition of the potential applications and prospects of spiking computing in three areas: model algorithms, sensing device integration, and neuromorphic hardware applications.

SpecForge: Open Source Framework for Training Speculative Decoding Models

Beat OpenAI and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning

This session explores the practical application of dataflow-oriented robotic architecture using Dora-rs, specifically focusing on humanoid robot development. The presentation will demonstrate how to build an AI robot from scratch using Dora's dataflow paradigm, covering the fundamental concepts of dataflow programming in robotics, implementation strategies for humanoid systems, and practical examples of sensor integration, decision-making processes, and actuator control within the Dora framework.

AIKG: AI Driven Kernel Generator

From Multimodal to Intelligent Agents: AI+AR Driving Human–Machine Symbiosis

Building on the Open Source Summer project work, this presentation shares experiences in integrating various external ecosystems with Dora and building unified launch processes around Dora. We explore the technical challenges of ecosystem integration, standardization of launch workflows, and best practices for creating seamless interoperability between Dora and other robotics frameworks. The talk covers practical implementation strategies, common integration pitfalls, and solutions for creating robust, unified development environments.

Group Photo

Overhauling Multimodal Support in llama.cpp

AIRSPEED: An Open-Source General-Purpose Data Generation Platform for Embodied Intelligence

MCP provides powerful capabilities for AI Agents, but practical MCP implementation faces many challenges. How to seamlessly integrate MCP Servers within enterprise, home, and mobile LANs for global access? How to quickly publish and deploy MCP Server tools? How to effectively manage numerous MCP Servers with flexible usage? How to set up on-demand access permissions for MCP Servers and tools with security? MCPdirect is an open-source MCP Universal Access Gateway from a Hong Kong startup team, designed to solve these critical integration and management challenges in the MCP ecosystem.

Intelligence Begins with Memory: Creating a New Paradigm for Memory Management Based on MemOS

Ventus is a high-performance open-source GPGPU based on RISC-V and its vector Extension to deliver a scalable, fully open-source GPU solution. By Professor Hu He from Tsinghua University, this project represents a groundbreaking initiative in creating accessible, transparent, and customizable GPU computing architectures. The presentation covers the technical foundations of the Ventus architecture, its RISC-V vector extension implementation, performance characteristics, and the broader implications for democratizing GPU computing through open-source hardware design.

Joint Algorithm-System Optimization Practices by the Qwen Team in Large Language Model Inference

Development and Exploration of Distributed Dora

Developing Applications on Linera with Rust: Microchain Programming Paradigm

Creation of an executable

3D Gaussian Splatting with Rust & WebGPU

This presentation will provide an in-depth introduction to the technical characteristics and ecological development of Dora-rs as a new generation of high-performance development framework for intelligent robots. The presentation will focus on showcasing Dora-rs's innovative practices in the open-source HarmonyOS ecosystem, including deep cooperation with universities, open-source HarmonyOS communities, and OpenLoong humanoid robots, as well as providing development software and hardware toolchains for the community. By sharing experiences from GOSIM hackathons, exhibitions, and community training courses, we envision how Dora-rs can promote the autonomous and controllable development of domestic robot technology through an open collaboration model, building an "AI-first" robot development ecosystem.

Exploration of Building a Commercial Data Warehouse Platform with Rust

The First Echo of Asynchrony: Sensing the Pulse of Scheduling from #[tokio::main]

Rust Applications on the Edge: Developing Scalable AI Frameworks for Embedded Devices

Bringing AI to the Edge with Dynamic Optimization

SpatialLM is a large language model designed to process 3D point cloud data and generate structured 3D scene understanding outputs. These outputs include architectural elements like walls, doors, windows, and furniture arrangements with precise spatial relationships. This presentation explores the novel approaches for training language models to understand spatial geometry, the challenges of bridging natural language processing with 3D spatial reasoning, and practical applications in robotics, AR/VR, and smart building management systems.

This presentation explores the current state and future directions of high-performance web engines, focusing on the latest developments in browser engine optimization, rendering performance, and computational efficiency. As Chair of the W3C High-Performance Web Engines Community Group, Martin will discuss emerging standards, best practices for engine development, and the challenges of balancing performance with web compatibility. The talk covers modern optimization techniques, GPU acceleration, memory management strategies, and the integration of AI capabilities into web engines to meet the demands of next-generation web applications.

Information Retrieval Beyond Keywords: Unlocking Regulatory and Standards Knowledge at Scale

Can Neural Networks Be Rigorously Explained? Will Scaling Laws Become an Eternal Constraint for Large Models?

The Huawei Ascend architecture differs significantly from NVIDIA GPU architecture, including differences in chip architecture, interconnect, software stack, programming system, and operator libraries, which brings many challenges for SGLang on Ascend adaptation and support. This presentation introduces the Ascend system architecture and shares the journey and efficient practices of SGLang on Ascend adaptation. We explore the technical hurdles overcome, performance optimizations achieved, and lessons learned in porting SGLang to the Ascend ecosystem.

Palpo Matrix Server

Implementation of Bidirectional Concurrent Linked List

Beyond Memory Safety: Challenges and Practices for Highly Available Rust Infrastructure

Hardware-Agnostic Acceleration of Open-Source Models with Hugging Face

Map Generation and Path Planning Approach for Unstructured Environments

OpenHarmony ArkWeb Progress and AI Exploration

We demonstrate how a custom RAG system without rerankers achieves strong internal MRR (Mean Reciprocal Rank) and powers our agents. Then we transform the same context-engineering approach into robot memory: short-term buffers, episodic logs, and long-term knowledge graphs. This unified architecture bridges information retrieval and robotic action planning, showing how memory-augmented systems can maintain contextual understanding across diverse AI applications. We explore the technical implementation, performance comparisons, and practical benefits of memory-centric approaches over traditional reranking methods.

The Curse of Depth in Large Language Models

Research on all-to-all Communication Optimization in Large-Scale Mixture of Experts (MoE) Models

The Art of Rust Type Abstraction: Thoughts and Practices on Building High-Productivity ORM

os-checker: Rust Code Analysis Platform with Automated Quality Monitoring

Rust + RDMA empowers cross-machine remote memory Serverless systems

Volcengine SGLang Best Practices - Deploying DeepSeek with PD Disaggregation and Expert Parallelism

This presentation introduces Qingcheng.ai's design and development of a training and inference optimization platform for indigenous large models, with improving indigenous intelligent computing power as the core application scenario. The platform adopts the self-developed Chitu inference engine, designs a unified framework for training and fine-tuning, and proposes hybrid precision and low-bit quantization optimization schemes. While maintaining computational accuracy and general problem-solving capabilities, it significantly improves model computational speed and reduces computational resource consumption through engineering and product practices.

vLLM is a fast and easy-to-use library for LLM inference and serving. In this talk, I will briefly introduce the evolution of the vLLM project, the open-source community behind it, and highlight some features that are interesting to many users.

Introduction to Ant Group's Ming-Omni series open-source work in the large model field, sharing practices and progress in multimodal large model architecture evolution, cross-modal fusion, and unified generation and understanding. Through joint design and optimization of model architecture and training processes, we are committed to building comprehensive modal capabilities for models, achieving multimodal foundation models that can see, hear, speak, and draw.

Unleashing GenAI Inference at Scale: A Deep Dive into the llm-d Project

Currently, SGLang exclusively supports the PyTorch AI framework and runs on Ascend NPUs through the torch_npu plugin. We designed a dual-framework collaborative solution for PyTorch and MindSpore, fully considering cross-framework resource reuse while simplifying interfaces to leverage MindSpore's performance advantages on Ascend NPUs. Developers can easily integrate MindSpore models into SGLang while maintaining unchanged user workflows. MindSpore's integration solution also provides a foundation for subsequent integration of other frameworks.

Beat OpenAI and Build the Best AI For You: Mobilizing First-Party Data with Fine-Tuning