AMD Advancing AI

Kunlun Zhu

Developer, OpenManus Project, OpenManus
Graduate Student Researcher, University of Illinois Urbana-Champaign

A beginner level, instructor-led, hands-on GPU workshop on setting up AI inference and developing agents on AMD GPUs. Learn from the founder of the OpenManus project and AMD architects how to develop an AI agent application on AMD Instinct™ MI300X GPU, using MCP, open-source models (Qwen3, Llama4), and the OpenManus framework in a browser-based workflow.

Preferred prerequisite: None 

Kunlun Zhu is currently a first-year MSCS student at UIUC. He has been a participant in many agentic open-source projects, such as OpenManus, XAgent, ChatDev, and ToolBench, receiving over 80k stars. He has published multiple works in ICML, ICLR, ACL, TMLR, and Nature Communications, receiving over 1200 citations about Agent and tool use. He was previously working as a Research scientist at Tsinghua University and intern at CMU, Mila, and one of the early founding members of ModelBest Inc. and OpenBMB.

Matthias Reso
Partner Engineer, AI, Meta

A beginner level, instructor-led, hands-on GPU workshop to teach Finetuning LLMs and Image Generation models with AMD GPUs. Learn from PyTorch and torchtune core contributors and AMD architects how to fine-tune LLMs and diffusion models with torchtune, explore benefits of finetune library like Unsloth, and work hands-on to experience the advantages of fine-tuned models.

Preferred prerequisite: Workshop 101 

Matthias is a Partner Engineer, AI at Meta with 12+ years of experience in computer vision and AI applications. As part of PyTorch's Applied AI team he loves to explore and overcome the boundaries of deep learning frameworks and help others to do the same. His interest span high-performance model optimization, distributed training, and inference.

Robert Shaw
Director of Engineering, Red Hat AI

An advanced user, instructor-led, hands-on GPU workshop equips you with tips and techniques to maximize the performance of at-scale vLLM deployment on AMD GPUs. Join this exciting course, led by a core maintainer of vLLM alongside AMD architects. Learn from one of the core maintainers of vLLM the challenges in production-scale model serving, optimize performance with deployable kernels, and apply them to an LLM to boost efficiency, gaining confidence in enhanced performance and TCO benefits.

Preferred prerequisite: Workshop 101

Robert Shaw serves as the Director of Engineering at Red Hat and has also held the position of Committer at vLLM. With a focus on engineering leadership and contributions to open-source projects, Robert Shaw demonstrates a strong commitment to driving innovation and quality in software development.

Mark Saroufim

Software Engineer, Meta
Co-Founder, GPU Mode

An advanced user, instructor-led, hands-on GPU workshop introduces AMD GPU architecture and guides you in developing optimized kernels. Join this exciting course, led by the co-founder of GPU Mode, alongside their team of core maintainers and AMD architects. Explore GPU architecture, learn kernel development basics from the founder of GPU Mode, one-of-its-kind community for GPU kernel development, start with a simple kernel example, and join a GPU Mode kernel project to build practical skills.

Preferred prerequisite: Workshop 201 

Mark Saroufim is a Software Engineer at Meta Platforms and core contributor to PyTorch, specializing in GPU optimization and performance engineering. He co-founded GPU MODE, an open-source research community with over 16,000 members dedicated to making GPU programming more accessible.

Daniel Han
CEO, Unsloth AI

Explore the latest trends to accelerate training and their impact on kernels, reinforcement learning, and quantization. Additionally, discover how to prepare for a post-training future shaped by reinforcement learning. Discover the technical foundations behind GRPO, the algorithm behind DeepSeek-R1, and unlock the power of reward functions.

Daniel Han is the CEO of Unsloth AI, an open-source AI startup on a mission to make AI accessible and accurate for everyone. Unsloth has grown rapidly, earning over 38,000 GitHub stars and surpassing 10 million monthly downloads on Hugging Face. The team has contributed critical bug fixes to major models, including LLaMA, Phi, Gemma, and Mistral. Before founding Unsloth, Daniel worked at NVIDIA, achieving a 2000x speedup in t-SNE visualization.

Wenlei Bao
Research Scientist, Bytedance Seed

Triton-distributed, an extension of the Triton compiler, optimizes computation, memory access, and communication for distributed AI systems. It supports overlapping techniques, outperforms hand-optimized code, and reduces development time compared to low-level programming.

Chris Lattner
CEO, Modular Inc
Co-Founder, LLVM

Explore using the Mojo programming and language and MAX framework to unlock the power of AMD GPUs for leading-edge Generation AI applications.

Chris Lattner is a co-founder and the CEO of Modular, which is building an innovative new developer platform for AI and heterogeneous compute. Modular develops MAX, an AI deployment system that accelerates GenAI inference, as well as the Mojo language - a Pythonic system for GPU programming and high performance. He co-founded the LLVM Compiler infrastructure project, the Clang C++ compiler, the Swift programming language, the MLIR compiler infrastructure, the CIRCT project, and has contributed to many other commercial and open source projects at Apple, Tesla, Google and SiFive.

Simon Mo
vLLM Project Co-Lead, vLLM

Explore overcoming LLM inference hurdles with vLLM, which provided memory-efficient methods and enables fast, scalable AI deployment for real-time applications across industries.

Simon Mo is a PhD student at Berkeley Sky Computing Lab. He is a co-lead of vLLM project, which focuses on building an end-to-end stack for LLM inference on your own infrastructure. The project aims to improve the throughput of popular LLMs by 2-4x with the same level of latency compared to other systems.

Joe Chau
VP of Engineering, HPC & AI, Microsoft Azure

Peng Cheng
Senior Principal Research Manager, Microsoft Research Asia

Discuss the technical complexities of scaling GPU-based training for modern AI models, including resource optimization, distributed computing bottlenecks, and energy efficiency. The session highlights innovative solutions and advancements in Azure’s infrastructure and research, offering insights into the future of high-performance computing for AI development.

Joe Chau is Vice President of Engineering at Microsoft, where he leads the Azure High Performance Computing (HPC) and Artificial Intelligence (AI) infrastructure team. He is responsible for building and operating the large-scale systems that power Microsoft’s most advanced AI workloads. Joe’s work spans cloud infrastructure, systems engineering, and hardware-software co-design, enabling the deployment of cutting-edge GPU clusters and next-generation compute platforms. He has played a key role in scaling Microsoft’s AI capabilities through collaboration with internal teams and industry partners. Known for his technical depth and strategic clarity, Joe helps ensure Azure remains a global leader in AI infrastructure.

Peng Cheng is a Senior Principal Research Manager at Microsoft Research Asia (Vancouver), specializing in systems, networking, and artificial intelligence. His recent work focuses on pioneering advancements in AI infrastructure and AI-system co-evolution. Peng has authored over 30 papers in leading systems, networking, and AI conferences, and his research has played a key role in enhancing Microsoft products, including Azure, Microsoft Teams, and Bing.

Lianmin Zheng
Member of Technical Staff, xAI
SGLang Project Lead

Discover SGLang, an efficient serving engine for large language models and vision-language models. Lianmin covers core inference optimizations and recent advancements, such as prefill-decode disaggregation, large-scale expert parallelism, zero-overhead scheduler, and speculative decoding. SGLang is widely adopted across the industry to serve frontier models like Grok3 and DeepSeek V3/R1.

Lianmin Zheng is a Member of Technical Staff at xAI. Lianmin co-leads the open-source development of the SGLang project and inference optimizations at xAI.