backend.ai

With the proliferation of generative AI, prohibitive development and adoption costs are becoming a major barrier to AI democratization. How can we lower the cost threshold and increase efficiency and ease of use when using or developing deep learning-based solutions? We introduce Slipstreaming, an approach from Rableup and Backend.AI to achieve massive performance and efficiency while standing on the shoulders of giants and masking all complexity.

About the presenter

- CEO, Lablup

- ML/DL Google Developers Expert

- Ph.D in Statistical Physics (complex systems / neuroscience) at POSTECH

A look back at the design philosophy and evolution of Backend.AI from an idea to provide a reproducible computational environment for researchers to an AI development and services platform for large enterprises. We explore why we made different choices than other similar products in the industry, and how those choices have delivered value to Lablup's business and customers.

About the Presenter

- Ph.D., Computer Science, KAIST (High Performance Network Systems)

- Open source ecosystem contributor including Python (CPython, aiomonitor, aiohttp, aiodocker, pyzmq, DPDK, iPuTTY, etc.)

Backend.AI's enterprise support services mean an ongoing partnership with you, not just a product installation. From scheduling the installation, we work closely with your sales and devops teams to provide customized installation and service optimization. After the installation is complete, we provide ongoing support that is responsive to customer feedback and needs, starting with user guide sessions. This presentation will provide an introduction to how we interact with Backend.AI enterprise customers.

About the presenter

- Director of Research, Lablup

- Ph.D., Department of Forensic Physics (Single Molecule Biophysics)

We will analyze the main problems with the existing WebUI and where the user experience needs to be improved. We'll discuss why the problems need to be solved, what needs to change and how to improve them, and how the UI/UX has changed, with a particular focus on the session launcher UI/UX. Finally, we'll make some predictions about how WebUIs will change in the future.

About the presenter

- Software Engineer, Lablup

Backend.AI's powerful giant AI model development technology is meeting GenAI to create the ultimate AI service capability. This session will introduce how the model serving capabilities introduced in Backend.AI 23.09 have been extended to the operational side of GenAI, in terms of technology, convenience, and fate resource savings. This will be a session that will inspire you with the many innovations that can be created by combining efficient operation of GPU resources with automation of GenAI services and fine-tuning.

About the presenter

-Software Engineer, Lablup

Lablup doesn't have a planner position, but we're still releasing a lot of products and new features, including Backend.AI. We'll talk about the many trials and errors that led us to use GitHub, a platform that developers are familiar with. In particular, we'll talk about how we ended up using GitHub Projects, how most of our development infrastructure is open source, when we tried to adopt Jira and why we stopped, and the underlying issues of scaling the company. We'll also talk about how we use the most important GitHub Projects, their shortcomings, and how we're using GitHub Actions and Labels to solve them.

About the presenter

- Software Engineer, Lablup

Lablup doesn't have a planner position, but we're still releasing a lot of products and new features, including Backend.AI. We'll talk about the many trials and errors that led us to use GitHub, a platform that developers are familiar with. We'll cover how we came to use GitHub Projects, how most of our development infrastructure is open source, when we tried to adopt Jira and why we stopped, and the scaling issues of our company. We'll also talk about how we use the most important GitHub Projects, their shortcomings, and how we use GitHub Actions and Labels to solve them.

About the presenter

- Software Engineer, Lablup
- GitHub Campus Expert

1. WASM, the Rise of WebGPUs
2. Performance evolution of edge devices - Apple M2 ultra beats A100
3. Agent configuration by sLLM
4. Federated Inference in the Age of Generative AI
5. Use cases for private generative AI - LLM + Diffusion

Speaker Introduction

I am the CEO of Sionic AI and previously led major AI projects at NAVER and TOSS. At NAVER, I led the development of AI products for Clova and NAVER Cloud, and developed a system to orchestrate tens of thousands of deep learning models. At Toss, he rebuilt the semantic search and brand recommendation systems for the main search service.

Currently, Sionic AI provides enterprise-class custom LLM services and is developing a SaaS product specialized in LLM Agent.

We are challenging the new generative AI market with our technology and business experience.

In this presentation, "Fine-tuning Whisper for Voice Novelty", I will introduce the painstaking process of fine-tuning the Speech-to-Text feature by utilizing Whisper, a model provided by OpenAI, for a team touching voice data for the first time.
In the age of LLM, he will also share how Whisper models and LLM can come together to transform new areas such as maritime communications, and how they leveraged Backend.AI to learn how to fine-tune Whisper models for their purposes.

About the Speaker

Currently, he is a professor at Pukyong National University, School of System Management Engineering, and has been conducting AI research mainly in the field of NLP. In addition to being a professor, he is also an entrepreneur who recently completed his LLM with ChatGPT and recklessly started an AI company in the field of trade, "TeamLibutte", believing that the world is changing due to the change of technology. As a company, I have to do voice fields other than NLP, so I am challenging it with tears (as a researcher).

Understand the background of the AI era sparked by generative AI and how to capitalize on it. In this session, you'll learn how real-world customers are utilizing LLM, how to prompt engineer with generative AI to meet business needs, how to choose the right LLM model, a comparison of RAG architecture and fine-tuning, how to integrate real-time APIs into LLM, building intelligent applications designed for a changed UX, and how to build complex applications as LLMOps.

About the presenter

Hyunsoo Kim is an expert technology strategist with development experience in various fields such as IoT, Back-end, Front-end, Automation, Data Analytics, AI/ML, and Cloud, and has worked with customers in various industries to build solutions that meet their needs. He is currently working as an expert technology strategist for digital native customers at Microsoft Korea, helping them solve their business needs using generative AI technologies. Learn more: http://www.studydev.com/me/

A collection of use cases from Korean customers utilizing Azure OpenAI:

https://www.youtube.com/playlist?list=PLGh_JNxzXsX9NSm-iyAdS4Ioco0vp4jtq

Hands-on with Auzre OpenAI: https://github.com/HyounsooKim/azure-openai-samples-kr

It's been a few years since I started running the TodayCode YouTube channel, and I've been asked to create content and teach a lot of different things, and I've taught a lot of different people in a lot of different places. I've taught students, job seekers, people who want to use data in their jobs, people who want to switch jobs, and I've talked to people who are struggling because they're non-majors, because they're older, because they have to work for a living right now, and they want to get into development or data, and they're struggling.

About the presenter

• Todaycode YouTube Channel Operator
• Microsoft MVP (Python Developer Technologies)
• Director of FastCampus Kernel 360
• Inflearn Knowledge Sharer
• NAVER Connect Boost CourseData Science Production and CoachingStudy Live Coach
• Lectures and content production for Seoul National University Big Data Innovation Sharing University, Yonsei University, Hanshin University, Korea Meteorological Administration, Statistics Korea, Seoul Digital Foundation, etc.
• Author of Everyone's Korean Text Analysis with Python

Discover the magic behind VisuTale AI, where images become the seeds of stories and art. We'll delve into the intricacies of each AI model, from interpreting images to crafting stories and generating related digital art. This session promises a deep dive into the models and engines that power VisuTale, offering a blueprint for those looking to harness the potential of generative AI on Backend.AI.

About the presenter

• Researcher, Lablup
• ESG Data Scientist, Who's Good
• MS, KAIST Graduate School of Data Science

High-speed storage is often used to train AI with large amounts of data or to quickly deploy AI models. Backend.AI supports a lot of high-speed storage, and I'd like to share how I worked on integrating vast data storage.

About the presenter

- Software Engineer, Lablup

We'll look at the accelerator plugin structure of Backend.AI, the accelerator plugins developed to date (Mock, CUDA, CUDA Enterprise, ROCm, TPU, IPU, ATOM, Warboy), additional accelerator plugins currently in development (SAPEON, HyperAccel), and discuss what needs to be done on the AI semiconductor side.

About the presenters

- Software Engineer, Lablup

Beyond AI - Jeongkyu Shin

Context

• We will share with many people the perspectives of developers, researchers, and platform company representatives about the changes in the deep learning field over the past two years and the changes that will occur over the next two years.

Speaker

• CEO, Lablup
• ML/DL Google Developers Expert
• POSTECH Doctor of Physics (Complex Systems Physics/Computational Brain Science)

I'm new to low-level Linux: A Python Developer's Journey into the Abyss (Kyujin Cho)  

Backend.AI supports accelerator interfaces like CUDA GPUs and Graphcore IPUs, as well as newer technologies like GPUDirect Storage. In this session, we'll share the shoveling we've been doing in Linux's system calls and networking layers to integrate these features.

Writing React components as if they were asynchronous - Jongeun Lee

Can you imagine frontend development without async? We'll walk through an example of what it means to write React components as if they were asynchronous, and talk about why this mindset clears up a lot of clutter in a developer's head.

FastAPI Struggles - Forklift (Kangmin Kwon)

I share my experience of working on an in-house project using FastAPI, an up-and-coming Python web framework. Introducing the struggles of the FastAPI project, which is still ongoing, through the eyes of a junior developer.

Writing machine learning - mostly write, little use (박해선)

The story of learning and exploring machine learning while writing a book.

About the presenters
- IT writer, translator
- GCP Champion Innovator
- Former ML GDE
- Author of 'Self-study Machine Learning + Deep Learning', 'Do It! An Introduction to Deep Learning', and translated several machine learning books into Korean, including 'Hands-on Machine Learning 2nd Edition' and 'Deep Learning from the Creator of Keras 2nd Edition'.

A Past Life in Open Source: One PR Blow (Sion Kang, Sang Hun Lee)

Description
- Xiong Kang: I would like to share tips from a novice developer who entered the open source ecosystem with a high barrier to entry. 
- Sanghoon Lee : He will share his experience as a mentor and mentee in the Open Source Contribution Academy and his experience as a junior developer interested in various open source projects.

About the presenters
- (Sion Kang) Intern at Lablup
- (Sanghoon Lee) Backend developer, Lablup

Debugging production issues with better visibility into asynchronous tasks (Joongi Kim)

Description.
- As you develop complex asynchronous apps with Python asynci, you may encounter issues like suddenly having too many of a particular task, or a task that should be running continuously dies silently. These issues are very difficult to debug using only postlogs or general tracebacks because they occur in the interaction of multiple tasks that are stacked separately from each other. I will share the process of developing the aiomonitor-ng library that can monitor the task creation and termination process in multiple stages, and the cases where we caught bugs with it.

About the presenter
- CTO, Lablup 
- Backend.AI Lead Developer
- CPython and aio-libs contributor
- PhD in Computer Science (Network Systems), KAIST

NVIDIA Omniverse for Generating Virtual Worlds(Dr. Pallavi Mohan)

NVIDIA Omniverse is an extensible platform for virtual collaboration and real-time, physically accurate simulation. Creators, designers, researchers, and engineers can connect tools, assets, and projects to collaborate in a shared virtual space. Creating digital twins in Omniverse allows you to create physically accurate virtual replicas of unique objects, processes, and environments, all constantly in sync with real-world data inputs and powered by AI. It can also a powerful tool for content generation and synthetic data generation. With Omniverse Replicator, an advanced and extensible SDK within the ecosystem, researchers, developers, and enterprises can generate physically accurate 3D synthetic data, and easily build custom synthetic data generation (SDG) tools to accelerate the training and accuracy of perception networks. Tune in to this talk to know more about the Omniverse ecosystem, and the possibilities of generating your own virtual worlds.

Speaker info.
- Solution Architect, NVIDIA AI Tech Center