PKU Weiming Supercomputing TeamLinux Club of Peking Unversity
How can we train faster, run better, and serve for less?
From kernels and compilers to distributed systems and collective communication, and from model serving to reinforcement learning systems, this seminar series connects the ideas and engineering behind large-scale AI training and inference.
Kernel & Operator
Start with GPU architecture and CUDA, then follow layout, data reuse, and pipelining through high-performance GEMM, Attention, and MoE kernels. Along the way, see how DSLs and ML compilers make performance engineering more productive and increasingly automatic.
Communication
Study interconnects and collective communication from multiple layers, and understand how they determine the real efficiency of large-scale AI systems.
Serving
Use KV cache, scheduling, and parallelism as the main threads for understanding inference optimization, then connect those ideas to the architecture of real serving systems.
RL Systems
When training, inference, rewards, and environments all run together, learn to reason about tail latency, resource scheduling, weight synchronization, fault tolerance, and training-inference consistency as one system.
Each seminar focuses on one systems topic and builds a structured understanding from fundamentals to practice.
Questions and follow-up discussions connect papers, source code, and real engineering experience.
Industry experts share current techniques, systems, and lessons learned in production.
Exercises, evaluation environments, and test cases cover CUDA, TileLang, training, and serving frameworks.
Slides, notes, and supporting materials are published openly and updated throughout the series.
