Schedule by date
October 17 Schedule
Join us for six exciting conference tracks, with full agenda, session times, topics, and speakers coming soon. Don' t miss out on these dynamic sessions! (More sessions continuously added)
October 17
Time
Session
17
8:30
Registration
Time
8:30
17
9:30
GOSIM Keynote: Open Source Ecosystem in the Era of AGI
Open Source Ecosystem in the era of AGI... More
Details Soon
Time
9:30
17
9:45
Opening Keynote: Mobile Solution by OpenHarmony
Huawei promotes the openness, concurrency, and
security of mobile devices operating systems based on OpenHarmony, and combines the
lightweight characteristics of devices to provide the industry with more effective
parallel and concurrency solutions.
Time
9:45
17
10:05
GOSIM Hackathon Awards Ceremony
Winner of the GOSIM Hackathon will be announced.
Time
10:05
17
10:10
Morning Break
20 Minutes Break
Time
10:10
17
10:30
Decentralized Technologies and Next-Generation Platforms
Digital experiences are heavily mediated by
centralized platforms. These dominant platforms face disruption by decentralized
building blocks, AI, and algorithms - affording new opportunities to rebuild our
platforms on open foundation, and reclaim control of our data and experiences. This
presentation shares insights gleaned from research interviews with 23 people working
toward that open future.
Time
10:30
17
10:30
OpenCV and Spatial Intelligence
I will describe the upcoming OpenCV release and
current initiatives into "
spatial intelligence"
(AI + 3D pose in a space)
Time
10:30
17
10:30
OPEN-SORA: Democratizing Efficient Video Production for All
The Open-Sora project is the world'
s first
open-source reproduction solution similar to Sora, creating an efficient and
low-cost video generation model. This presentation will begin by introducing the
core concepts of Open-Sora, providing an in-depth analysis of its unique model
architecture, and demonstrating how this architecture optimizes video generation
efficiency while significantly reducing costs. The presentation will also discuss
how Open-Sora adapts to different video resolutions and lengths while maintaining
high-quality output, and how it addresses common data storage challenges in large
model training. Additionally, it will cover optimizations in inference cost and
memory usage. Open-Sora has received enthusiastic responses in the open-source
community due to its fun and high-quality results. In this presentation, we will
also showcase creative works from the open-source community and explore cutting-edge
video generation technologies with you.
Time
10:30
17
10:30
WGPU: Modern GPU Programming In Rust
GPUs have great performance—if you can make your
problem fit their computational model. Hyperthreading, local memory spaces,
mandatory vectorization, and other quirks all affect how GPU code should be written.
WGPU is a cross-platform API for GPU coding.
Time
10:30
17
10:30
Next Generation App Development with Dioxus: Fullstack,
Cross-Platform, Strongly Typed, and Blazingly Fast
Dioxus is a Rust framework that makes it makes
it easy for developers to ship web, desktop, and mobile apps with a single unified
codebase. In this talk, we’ll be unveiling features like native mobile support, live
hot-reloading, and a custom AI model.
Time
10:30
17
10:30
Register Allocation in Cranelift
Register allocation is one of the most important
stages in an optimizing compiler and is often responsible for a large fraction of
the compilation time. This talk presents the register allocator used by Cranelift, a
fast compiler backend written in Rust and used by the Rust compiler, with a focus on
regalloc3, its newest iteration that is currently being developed.
Time
10:30
17
10:30
Makepad in Action: Building your First Application
This hands-on workshop introduces the power and
flexibility of Makepad, a cutting-edge UI rust framework for cross-platform
development. Participants will: Learn the core concepts of Makepad'
s GUI
Framework
Build a functional application from simple widgets to advanced features
Implement modern UI elements and interactions
Gain practical experience applicable to both mobile and desktop platforms Suitable
for developers of all levels, from beginners to experienced programmers. By the end
of the workshop, you'
ll have created your own cross-platform application and
acquired the skills to continue developing with Makepad.
Time
10:30
17
10:30
Workshop: W3C DID Forum
W3C'
s Decentralized Identity specification
is designed to enable individuals and organizations to have greater control over
their personal data and online identities. By using decentralized identifiers
(DIDs), users can create and manage their identities independent of centralized
authorities, offering a more secure and privacy-enhanced approach to digital
interactions. DID has been one of the milestone technologies for the Next Generation
Internet.
W3C DID Forum is proud to feature Mr. Drummond Reed, the co-editor of the W3C DID
1.0 specification talking to us the initiation, progress and implementation story of
DID, with his vision how DID will reshape the digital world. Welcome to register and
join us on the journey.
Drummond has spent over a quarter-century in Internet identity, security, privacy,
and trust infrastructure. He is Director of Trust Services at Gen (formerly Avast)
after their acquisition of Evernym, where he was Chief Trust Officer. He is
co-author of the book, Self-Sovereign Identity (Manning Publications, 2021). He is
founding board member of the OpenWallet Foundation and the Trust Over IP (ToIP)
Foundation, where he serves as co-chair of the Technology Stack Working Group and
the Concepts and Terminology Working Group. In 2002 he received the Digital Identity
Pioneer Award from Digital ID World, and in 2013 he was cited as an OASIS
Distinguished Contributor.
Time
10:30
17
10:30
Contribution-Oriented Open Source Talent Evaluation System
and Its Application
The contribution-based open source talent
evaluation system provides a fair and transparent incentive mechanism for open
source communities while helping companies identify and cultivate open source
talent. This presentation introduces the open source governance standard released by
the Ministry of Industry and Information Technology, which defines the measurement
of open source contributions. Through specific case studies, the application of this
evaluation system in practice is demonstrated. By co-creating the OpenTalent open
source talent evaluation platform, we aim to assist enterprises, educational
institutions, and governments in accurately identifying and cultivating open source
talent, thereby promoting the healthy and sustainable development of the global open
source ecosystem.
Time
10:30
17
10:45
How Universities Can Participate in Open Source
Contributions
How Universities Can Participate in Open Source
Contributions in OSPP
Time
10:45
17
11:10
Eclipse Zenoh: the Communication Backbone for Ubiquitous and
Heterogeneous AI Use-Cases
AI has become ubiquitous, enhancing many of our
interactions with new technology. This omnipresence requires a powerful
communication backbone allowing data to flow seamlessly. This presentation details
Eclipse Zenoh, a next-gen communication middleware.
Time
11:10
17
11:10
Building for a User-Centered Internet
The Internet is increasingly centralised around
a few big platforms, leading to numerous problems ignoring or harming users'
benefits. In recent years, more attention is paid on decentralised architectures
with more user focus. In this talk, I will briefly revisit the different lines of
thinking, and focus on one promising approach, Solid (Social Linked Data), proposed
by Web'
s inventor, Tim Berners-Lee, to re-decentralised the Web. I will cover
its key principles, what it offers right now, and (our team'
s) explorations to
expand its boarder. The talk will close with visions upon Solid and corresponding
open questions to incentivise future evolution.
Time
11:10
17
11:10
Through the Eyes of a Maintainer: How to Make Your
Contribution More Successful
Whether its to the Rust Project, your favorite
Rust tool, or a library you use, your first contribution can be intimidating. You
are stepping into an area where you don'
t know the social dynamics or technical
expectations. I'
ll be stepping you through the process of contributing from the
maintainer'
s perspective so you can better understand how this works and how to
make it more likely your contribution is both accepted and quickly.
Time
11:10
17
11:10
Construction and Open-Sourcing of the Bilingual Large Models
in the Aquila Series within the FlagOpen
This presentation primarily introduces the
construction and open-sourcing of the bilingual large models in the Aquila series,
specifically including model pretraining, open-sourcing of the training framework,
and open-sourcing of the datasets.
Time
11:10
17
11:10
Flutter: Google’s UI toolkit for Mobile, Web, &
Desktop
Apps from a Single Codebase
Flutter is an open source framework for building
beautiful, natively compiled, multi-platform applications from a single codebase.
Time
11:10
17
11:10
Introduction to the Latest Progress of the Khronos Vulkan
Standard
Introduces the Khronos Vulkan standard, the
latest progress of the standard, and some new Vulkan features.
Time
11:10
17
11:50
On the Role of Open-Source Simulation in the Era of Embodied
AI
In the era of Embodied AI, open-source
simulation platforms play a pivotal role in driving innovation and collaboration.
This talk explores the balance between openness and closed software/APIs,
emphasizing the need for flexibility and extensibility in developing robust
simulation tools for Embodied AI. By examining CARLA as a case study, I highlight
how open-source initiatives can foster community engagement, accelerate research,
and provide adaptable tools that meet diverse research needs.
Time
11:50
17
11:50
WGMath: Reusable WebGPU Shaders for Cross-Platform GPU
Scientific Computing
Using WebGpu, WGMath aims to bring GPU
scientific computing to all platforms by empowering all scientific communities with
reusable WGSL shaders. We discuss the fundamentals of this ecosystem, current and
future features for LLM, geometry, and physics.
Time
11:50
17
11:50
Rust Compiler Overview
This talk will provide an overview of the Rust
compiler: what it looks like, how it works, how its internals are structured, and
how it is developed.
Time
11:50
17
11:50
Blitz: Blurring the Lines Between Web and Native
Introducing Blitz, a new cross-platform
application runtime which combines web-compatible "
write-once, run-everywhere
"
UI code with a small footprint and powerful system integrations when
deploying to native (non-web) platforms.
Time
11:50
17
11:50
Palpo Social Server: A self-hosted social server supporting
Matrix protocol
A Rust based self-hosted, open-source Matrix
server for personal cloud deployment.
Time
11:50
17
11:50
InternLM: An Open-Source LLM with Full-Stack Toolchain
InternLM is developed by Shanghai AI Laboratory.
We keep open-sourcing high quality LLMs as well as a full-stack toolchain for
development and application. In this talk, we will share the latest developments of
InternLM and its related toolchain, covering areas such as Agents, Evaluation, and
AI-Infra.
Time
11:50
17
12:30
Lunch Break
Time
12:30
17
14:00
Writing Your Own Rust Linter
This talk will show you how to write your own
rust linter using the rustc_tools crate which allows you to have access to the rust
compiler API. It can be quite useful if you need specific lints for a project and
cannot add them to clippy.
It will also explain how the compiler API works and give a small example at the end
showing it in action.
Time
14:00
17
14:00
Open-Sourced Text-to-Video Model: CogVideoX
We introduce CogVideoX, a large-scale diffusion
transformer model designed for generating videos based on text prompts. Results show
that CogVideoX demonstrates state-of-the-art performance across both multiple
machine metrics and human evaluations. The model weight of CogVideoX is publicly
available at https://github.com/THUDM/CogVideo.
Time
14:00
17
14:00
Build an Open Source Operating System for Robots Based on
OpenHarmony.
Technological progress, an aging population and
many other factors are driving the booming development of the robotics industry, but
there is currently no official operating system that is fully oriented to robots on
the market. OpenHarmony is an operating system for smart terminal devices that is
oriented to the full-scenario, fully connected and fully intelligent era. Its
technical features such as the distributed capabilities of the soft bus are highly
consistent with the robot'
s demands for an OS. Shenzhen Kaihong, in conjunction
with a number of leading companies and universities, launched an open source version
of the robot operating system based on OpenHarmony in the open source community
under the guidance of the OpenHarmony Technical Steering Committee. This topic will
focus on the system architecture, technical features, industry applications and
other aspects of this OS.
Time
14:00
17
14:00
From Blockchain to Web3: Building the Decentralized Network
Scalability has become a key research direction
at the bottom of blockchain. The academic community has proposed several
representative solutions to improve the scalability of blockchain, such as directed
acyclic graph technology, sharding technology, new consensus mechanism, state
channel, side chain, cross-chain and other solutions. Among them, sharding
technology is considered to be a Layer 1 expansion solution with great potential to
greatly improve the scalability of blockchain. This report first sorted out the
research status of blockchain sharding technology, and then introduced the new DeFi
protocol built by Professor Huang'
s research team based on sharded blockchain -
BrokerFi.
Time
14:00
17
14:00
Simulation, testing, verification, and validation of
autonomous driving
SIMULATION, TESTING, VERIFICATION, AND
VALIDATION(STV2) is a set of processes that support the development, validation, and
operation of autonomous driving systems from the perspectives of safety and cost.
The scope, architecture, and critical components of STV2, as well as how the full
lifecycle of autonomous driving systems is covered in this talk. STV2 is part of the
infrastructure and tooling layer of autonomous driving architecture.
Time
14:00
17
14:00
Demystifying the LLM Training with the Fully Open-Source
Moxin 7B Model
Recently, Large Language Models (LLMs) have
undergone a significant transformation, marked by a rapid rise in both their
popularity and capabilities. Open-source LLMs, such as LLaMA and Mistral, have made
great contributions to the ever-increasing popularity of LLMs due to the ease to
customize and deploy the models across various applications. Although LLMs offer
unprecedented opportunities for research and innovation, its commercialization has
raised concerns about transparency, reproducibility, and safety. Many open LLM
models lack the necessary components (such as training code and data) for full
understanding and reproducibility, and some use restrictive licenses whilst claiming
to be “open-source”, which may hinder further innovations on LLMs. To mitigate this
issue, we follow the Model Openness Framework (MOF), a ranked classification system
that rates machine learning models based on their completeness and openness,
following principles of open science, open source, open data, and open access. We
present a truly open source LLM Moxin 7B and release pre-training code and
configurations, training and fine-tuning data, and intermediate and final
checkpoints, aiming to make continuous commitments to fully open-source LLMs.
Time
14:00
17
14:00
Triton Language &
vLLM Workshop Opening Remarks
Opening Remarks
Time
14:00
17
14:00
A New Generation of Programming Language for All-Scenario
Applications - Cangjie
An overall introduction to Cangjie, a new
generation of programming language for all-scenario applications
Time
14:00
17
14:00
uni-app x: The next generation, pure native cross-platform
framework
In recent years, cross-platform development has
long been a familiar development paradigm, but cross-platform development is still
often criticized by developers? Why? How to improve the solution?
In this speech, I will share the thinking and exploration of uni-app team on this
problem, and expound their performance bottlenecks and optimization practices for
the three platforms of native, H5 and mini program, and introduce uni-app x, a new
cross-platform framework. uni-app x uses a TS-like DSL, compiles to
swift/Kotlin/arkTS through Rust, and obtains a pure native App. On the premise of
completely solving the performance problems of traditional Hybrid apps, Uni-App X
uses the web ecology to liberate productivity and deliver quickly.
Time
14:00
17
14:00
The Dexterous Hand of Self-Evolution，The Key to Opening the
Door to Embodied Intelligence
The dexterous hand of self-evolution，The key to
opening the door to embodied intelligence
Time
14:00
17
14:25
Introduction of Key Technology of Cangjie Assisting in
Efficient Hongmeng Native Application Development
Introducing the key technologies involved in
Cangjie Hongmeng application development
Time
14:25
17
14:30
vLLM in Moonshot
1.The large-scale application of vLLM within
Moonshot
2.vLLM and the Mooncake PD separated architecture
3.Prospects for future development directions
Time
14:30
17
14:30
Introduction to ChinaEV100
ChinaEV100 is a conference platform for
interdisciplinary, cross industry, cross departmental, cross ownership, unofficial,
and non-profit policy and academic research in the field of electric and intelligent
vehicles in China;
A high-end exchange platform voluntarily participated by
government officials, experts, scholars, and industry professionals;
It is a
third-party think tank of the country in the field of automobiles.
Time
14:30
17
14:40
Incremental Linking and Hot Code Reloading for Rust
Slow edit-build-run dev cycles are a common
complaint among Rust developers. The Wild project aims to help by building a linker
in Rust with the goal of eventually supporting incremental linking and hot code
reloading.
Time
14:40
17
14:40
Life as a Rust Project Contributor
Join Jane Losare-Lusby as she takes you on a
journey through her experiences as a contributor to the Rust project. Starting with
her early work on Clippy, Jane has navigated various roles, including error-handling
project group lead, member of the Library and Style teams, and as a board member of
the Rust Foundation. After a brief hiatus to work on Truffle and June, she returned
to contribute to the Rust Compiler team, where she is now focused on the development
of locally stored metrics.
In this talk, Jane will share personal insights on navigating challenges such as
impostor syndrome, conflict resolution, burnout, and isolation. Jane will offer
practical advice for newcomers looking to contribute to Rust, highlighting the
importance of mentorship and connection, the dynamics of various Rust teams, and how
to turn an idea—like her current metrics initiative—into a meaningful contribution.
Whether you'
re a seasoned developer or just getting started, this talk will
provide valuable perspectives on how to find your place and make an impact in the
Rust community.
Time
14:40
17
14:40
Task Modeling Method for Functional Unmanned Vehicle
As a vehicle for smart missions such as
delivery, selling, shuttling and security patrol, functional unmanned vehicles can
be regarded as typical representatives of embodied intelligence. Considering of
variety of smart tasks, designing a set of general method for task modeling and
performing is of great significance to achieve the rapid development and deployment
of functional unmanned vehicles.
Time
14:40
17
14:40
Robrix: a Pure Rust Multi-Platform Matrix Client and More
Robrix is a new Matrix chat client that aims to
be a hub for federated services. By building atop the Robius app dev framework and
the Makepad UI toolkit, Robrix works seamlessly across multiple platforms (including
mobile) with no platform-specific code.
Time
14:40
17
14:40
Towards Practical Multimodal Large Models
As the inevitable path towards general
artificial intelligence, multimodal large models have shown great potential for
intelligent transition. It is not only at the forefront of academic exploration but
also a catalyst for building a community with a shared future for mankind and
promoting global cooperation. High deployment and inference costs, frequent
hallucination behaviors, and scarce high-quality data all greatly restrict the
development of multimodal large models. Starting from these key issues, MiniCPM-V
has for the first time achieved comparable end-side multimodal understanding
capabilities to closed-source commercial multimodal large models.
Time
14:40
17
14:40
Open-Sora Plan: An open source plan for large-scale video
generation models based on domestic computing power
Open-Sora Plan is an open source plan for large
domestic video generation models that supports the entire calculation on Ascend
computing power. This speech introduces the progress, details and planning of
Open-Sora Plan.
Time
14:40
17
14:50
Cangjie'
s Development Practice in Native Applications
for Hongmeng
This article introduces how to use Cangjie to
develop Hongmeng native applications, and provides a reference for developers to
develop Cangjie version of Hongmeng applications.
Time
14:50
17
15:00
vLLM Meets Qwen: What Have We Done in Alibaba?
The talk will share insights into the
optimizations made within Alibaba'
s Tongyi Lab to support the efficient service
of the Qwen model on vLLM. These optimizations encompass model quantization,
framework performance enhancements, and algorithmic improvements. Additionally, the
presentation will highlight the team'
s contributions to the vLLM open-source
community and conclude with a discussion on potential new challenges that the Qwen
series of models may pose to vLLM in the future.
Time
15:00
17
15:00
Afternoon Break
30 Minutes Break (SDV Workshop only)
Time
15:00
17
15:15
Cangjie Language Open Source Microservice Framework
Development Practice Sharing
Cangjie language, with its unique syntax and
powerful functions, has brought new vitality to the field of programming. Combining
the flexibility and scalability of microservice architecture, while exploring the
boundaries of modern software development, an open source microservice framework
based on Cangjie language was open sourced.
This project aims to combine the characteristics of Cangjie language with the
advantages of microservice architecture to provide developers with an efficient,
reliable and easy-to-maintain solution, whether it is a highly customizable Internet
application or an enterprise-level solution that requires stable operation, hoping
to create a development environment that can adapt to rapidly changing needs.
Time
15:15
17
15:20
Afternoon Break
30 Minutes Break
Time
15:20
17
15:30
Scenario Representations for Autonomous Driving Simulation
and the Oasis Simulation System
The standardization effort on scenario
representations for autonomous driving simulation will be introduced. Various
elements pertaining to a simulation system will be discussed, e.g., sensor models,
vehicle dynamics models and agent behavior models, etc., while taking as example the
Oasis Simulation System.
Time
15:30
17
15:30
vLLM: A Simple, Efficient, and User-Friendly Framework for
Large Model Inference
This presentation will cover the following key
points:
1.The inception and evolution of the vLLM project.
2.Models, hardware, and features supported by vLLM.
3.The community building of vLLM and an outlook on recent plans.
Time
15:30
17
15:50
Fireside Chat
A cozy interview and Q&
A with Dr. Josh
Triplett, from the Rust Project language, library, and Cargo teams.
Time
15:50
17
15:50
CodeGeeX4: A Versatile Code Generation Model for AI
Programming
We introduce CodeGeeX4-ALL-9B, the open-source
version of the latest CodeGeeX4 model series. It is a multilingual code generation
model continually trained on the GLM-4-9B, significantly enhancing its code
generation capabilities. Using a single CodeGeeX4-ALL-9B model, it can support
comprehensive functions such as code completion and generation, code interpreter,
web search, function call, and repository-level code Q&
A, covering various
scenarios of software development. CodeGeeX4-ALL-9B has achieved highly competitive
performance on public benchmarks, such as BigCodeBench and NaturalCodeBench. It is
currently the most powerful code generation model with less than 10B parameters,
even surpassing much larger general-purpose models, achieving the best balance in
terms of inference speed and model performance.
Time
15:50
17
15:50
Porting Servo to OpenHarmony
A progress update on porting Servo, a web
rendering engine written in Rust, to OpenHarmony
Time
15:50
17
15:50
The Development and Evolution of FaceChain in the Digital
Human Domain
FaceChain is an innovative open-source project
in the field of digital humans, leading the new wave with advancements in facial
perception and understanding, digital portrait generation, and talking video
creation. With just a single photo, users can create personalized digital avatars
and generate unique portrait works using a diverse range of styles and template
libraries. FaceChain offers a full-stack digital portrait generation capability,
including a 10-second training-free digital portrait generation feature and a DIY
style model quick training function, and is currently building an ecosystem for
digital portrait content. In addition, FaceChain has also laid out plans and
produced outputs for generating talking videos and virtual try-on experiences. With
its open-source spirit and vibrant community, FaceChain has driven rapid
technological development and widespread application. This topic will delve into the
technical features of FaceChain, community building, commercial application
prospects, and its evolutionary path in the digital human domain.
Time
15:50
17
15:50
Robot technology driven by large models in the industrial
field
In the industrial field, there is a growing
demand for data-driven robotics. The technical paths include advanced sensor tech,
machine learning algorithms, and real-time data processing. However, we also face
challenges like data security and interoperability.
Time
15:50
17
15:50
Moxin: Enhancing the Decentralized Distribution Capabilities
of Large Models
With the development of artificial intelligence,
large language models (LLMs) are increasingly used in various industries, but the
centralized distribution of model files still faces problems such as high
infrastructure costs and limited access. Moxin aims to solve these challenges by
distributing model files in a decentralized way and focusing on efficient and secure
transmission of model files.
Time
15:50
17
15:50
ICBC Hongmeng Native Application Cangjie Development
Practice
Share ICBC'
s practical experience in
developing Cangjie, the native application of Hongmeng, from the aspects of pilot
scenarios and implementation plans.
Time
15:50
17
16:00
The Open and Collaborative Innovation in automotive
operating system empowers the innovation and development of the automotive industry
Amid the wave of technological innovation in the
automotive industry, the secure and controllable automotive operating systems has
become central. Yet, the complexity of the technology, coupled with substantial
investments and lengthy development cycles, has created a rift between skilled
personnel and technological advancements. Addressing industry issues, new OEM
requirements, and international trends and policies, an open-source development path
is a reasonable approach. This presentation will concentrate on open-source
empowerment for automotive OS, the iSOFT open-source projects, and the collaborative
development plans.
Time
16:00
17
16:00
Implementation of the Triton Compiler Based on Moore
Threads'
Full-Function GPU
1.Introduction to the full-function GPU
architecture of Moore Threads.
2.Overview of the MUSA software ecosystem by Moore Threads.
3.Implementation of the Moore Threads Triton compiler and adaptation of FlagGems.
Time
16:00
17
16:10
Cangjie Development Practice for Leetcode Hongmeng Native
Application
Share the practice of developing Leetcode
applications on Hongmeng using the Cangjie language.
Time
16:10
17
16:30
Compression Carcinized: Drop-in Compatible Compression in
Rust
Data compression is used when loading almost
anything on the web. This talk is about how we implement zlib-rs and other
compression algorithms in rust so that they are drop-in compatible with their C
equivalent.
Time
16:30
17
16:30
Low-Level: Rust Education'
s "
Next Big Thing"
The Rust community has made a good start at
teaching tools for introductory Rust. I will review the state of Rust education,
especially in Universities, and talk about ways in which Rust-Edu and other groups
hope to fill new needs in the low-level space.
Time
16:30
17
16:30
What Data Imperative for Action Learning of Embodied AI?
After occurring of large scale language models
(LLMs), the development of Embodied AI is seen as a fundamental path to achieving
artificial general intelligence (AGI). However, data is currently a key bottleneck
for advancement in Embodied AI, compared to that of LLMs. In this talk, the learning
strategies are first analyzed, especially for embodied manipulation. Meanwhile,
various existing public datasets in embodied AI are investigated. Eventually,
requirements for an imperative embodied AI dataset are concluded and an upcoming
synthesized video dataset (rendered by Unreal 5) called MVGameIR is introduced.
Time
16:30
17
16:30
Functionary: Using tools with LLMs
Functionary is an open-source LLM that is
capable of using tools while maintaining conversational abilities. We will discuss
how complex agentic tasks can be achieved with function calls and how we addressed
challenges of function calling.
Time
16:30
17
16:30
GenUI: Declarative Rust Cross-platform Framework Based on
Makepad
GenUI is an innovative SFP front-end framework
developed in the Rust language, was originally inspired by Vue3 and Makepad.
Intended to help users write front-end projects using Rust more efficiently.
Time
16:30
17
16:30
Pointrix: a differentiable point-based rendering libraries
Over the past year, significant progress has
been made in differentiable point cloud rendering techniques, represented by 3D
Gaussian splatting. This report will introduce Pointrix, a differentiable point
cloud rendering optimization framework developed by our research group. Pointrix
consists of a modular Python API, an efficient CUDA backend, and a user-friendly
real-time rendering GUI interface, making it an open-source framework for point
cloud rendering that is easy to extend. Pointrix offers detailed quick-start guides
and rich example configurations, allowing users to implement specific algorithms by
making changes at the Python level. It provides researchers with a portable
environment for secondary development. Additionally, Pointrix supports a range of
mainstream 3DGS-related work, offering strong support for research and applications
in the field of differentiable point cloud rendering.
Time
16:30
17
16:30
CARLA Open Source Project Updates
CARLA New Release &
Features. CARLA is an
open-source simulator for autonomous driving research. CARLA has been developed from
the ground up to support development, training, and validation of autonomous driving
systems.
Time
16:30
17
16:30
Multi-Modal Compression and Inference with vLLM: Exploring
FlagScale'
s Application Practices and Technical Aspects
Large models have garnered significant attention
due to their exceptional performance across various tasks. However, in
resource-constrained scenarios, the substantial computational and memory resources
required for inference present numerous challenges. As a result, the industry is
actively developing technologies to enhance the inference efficiency of large
models. This report will share FlagScale'
s practical experience in compressing
and inferring multi-modal large models based on the vLLM framework: 1) a focus on
analyzing the related modules, strategies, and performance of the vLLM framework
with the newly added CFG Sampling feature;
2) using the llm-compressor tool to
perform quantization compression of multi-modal models at varying granularities
according to different deployment scenarios, while also exploring the differences
between multi-modal models and language models, as well as how to achieve extreme
compression of multi-modal models.
Time
16:30
17
16:30
Sharing Development of Enhanced Multi-Factor Identity
Authentication Components in Hongmeng TEE Environment PKI Architecture
Cangjie Programming Language + Hongmeng
Operating System Case Sharing
Time
16:30
17
16:50
Cangjie Development Practice of Weaver Software'
s
Hongmeng Native Application: EMobile10
Introducing the development of the Weaver
Software'
s Hongmeng native application EMobile10 Cangjie and sharing the
experience of using Cangjie
Time
16:50
17
17:00
Triton China Community Developer Contribution Program
Triton China Community Developer Contribution
Program
Time
17:00
17
17:10
Typical Use of Rust Programs in Different Linking Ways in
Trading Systems
This talk will delve into the fundamental
concepts of workspace, target, and crate-type. It will explore three cases within a
trading system, highlighting code dependencies and static links using
-crate-type=lib, dynamic link with -crate-type=cdylib, and static link using
-crate-type=rlib. Furthermore, the talk will expand on important aspects such as the
performance differences between dynamic and static links and the trade-offs in terms
of toolchain version flexibility when using rlib.
Time
17:10
17
17:10
Practical Exploration of XVERSE Large Language Model Series
This presentation primarily showcases the
hands-on exploration of developing the XVERSE large language model series, covering
various stages from ideation to realization and evolving from dense models to
Mixture-of-Experts (MoE) models. In terms of data, we will outline the necessary
datasets for effective training, describe preprocessing protocols, and discuss
approaches for continual model improvement. Regarding the model aspect, there will
be a focus on adapting from standard dense layers to implementing MoE structures,
highlighting decision criteria for specifying expert capacities and allocating
importance among them. Furthermore, addressing architectural facets, we will
articulate methodologies designed to maximize training productivity and secure
consistent learning performance.
Time
17:10
17
17:10
Cangjie for InfoQ
InfoQ Cangjie Practice
Time
17:10
17
17:10
Open World Embodied Large Models
Multimodal Large Language Models (MLLMs) have
demonstrated potential in visual instruction following across various tasks.
Recently, some studies have integrated MLLMs into robotic manipulation, allowing
robots to interpret multimodal information and predict low-level actions. While
MLLM-based policies have shown promising progress, they may predict failure
execution poses when faced with novel tasks or categories. Given these challenges,
we raise a question: “Can we develop an end-to-end robotic agent that not only
possesses manipulation skills but also effectively corrects low-level failure
actions?” Drawing inspiration from Daniel Kahneman'
s assertion that "
human
thinking is divided into a fast system and a slow system, which separately represent
intuitive processes and more logical reasoning,"
we introduce a series of
research works that mimic a human-like thinking paradigm to address the above
question.
Time
17:10
17
17:30
Patterns in Optimizing Triton Kernels for Reduction Ops
Triton language offers a tile-oriented,
CTA-level programming paradigm, attaining a good balance between control of hardware
and mental burden, which allows for non-gpu experts to author kernels with
reasonable performance within short time. But when writing triton kernels for a
library of general purpose, we need to choose appropriate algorithms and tasking
partitioning scheme according to the size of the problem or the layout of the data
to achieve better performance. This report shares some common tricks and patterns
for optimizing reduction-like kernels with softmax as an example, including
persistent reduction, online softmax normalizer, split-reduction and tasking
partitioning scheme for outer-reduction.
Time
17:30