AI Infra Seminars Wiki

Curriculum

The seminar moves from a single GPU kernel to the system architecture behind large-scale model training, inference, and reinforcement learning.

Topic	Weeks	Main subjects
Kernel and ML Compilers	1–3	GPU programming, memory hierarchy, Tensor Cores and layouts, software pipelines, GPU DSLs, ML compilers, and state-of-the-art kernels
Interconnect and Communication	4	Server interconnects, network fabrics, collective communication, communication libraries, overlap, and memory-storage co-design
LLM Serving and Inference	5–6	Metrics and benchmarks, prefill-decode disaggregation, KV-cache-centric systems, batching, inference kernels, speculative decoding, parallelism, and serving frameworks
Distributed Reinforcement Learning Systems	7	RL algorithms and frameworks, scheduling and resource allocation, asynchronous execution, rollout acceleration, training-inference consistency, fault tolerance, environments, and sandboxes

Guest lectures complement each topic with perspectives from current research and production systems.