See LionPATH to schedule.
Course Topic: Secure and Efficient Inference for LLMs & Agents
Inference systems are the "real-world interface" of LLMs: they determine latency, cost, reliability, and what information can leak or be manipulated when models are deployed at scale. This course studies inference-time (deployment-time) efficiency and security for LLMs and LLM-based agents. We will cover the modern LLM serving stack (prefill vs. decode, batching/scheduling), core efficiency techniques (KV cache, decoding algorithms, quantization), and the emerging security risks that arise specifically at inference time (routing integrity, cache-enabled leakage, cost-amplification/DoS, prompt injection in tool-using agents, and memory/RAG poisoning). The course is paper-driven: students read and review papers, lead a lecture-style presentation, and complete a group project.
