CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026)
by UC San Diego (Yiying Zhang)
A graduate systems course on what runs under LLMs and agents: inference, KV caches, GPU kernels, edge NPUs, serving, agent sandboxes and RL infrastructure.
Overview
CSE 291A: Systems for LLMs and AI Agents is a Fall 2026 graduate course at UC San Diego taught by Yiying Zhang, an associate professor of computer science and engineering who runs the WukLab and MLSys@WukLab research groups and founded GenseeAI. Her lab publishes on LLM serving, computer-use agents, RL post-training and GPU programming, and her awards include an OSDI Best Paper and an NSF CAREER award. The course page says it is not about ML theory, agent frameworks or vibe coding; it covers the backend. The public schedule has 17 lectures, each with linked PDF slides and one primary paper plus additional readings. It starts with agent foundations (ReAct, Anthropic's Building Effective Agents), then LLM inference (prefill and decode, Orca, PagedAttention), GPU kernel programming (ThunderKittens), and two lectures on edge AI and NPUs (PowerInfer, heterogeneous mobile SoC inference). The middle covers agent tooling and memory (MemGPT), OS support for agents, agent security (a USENIX Security '26 SoK), coding and computer-use harnesses (OSWorld-Human), multimodal agents and retrieval infrastructure (METIS). It ends with distributed serving (DistServe, MegaScale-Infer), agent serving (Agentix, NSDI '26), distributed training (PipeDream) and RL fine-tuning infrastructure (HybridFlow/verl). Enrolled students write paper summaries and do a research project worth 75% of the grade; there is no homework or exam. A few readings, such as TClone and OSWorld-Human, come from the instructor's own lab.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 10 weeks, 17 lectures of 80 minutes; self-study via public slides plus ~1 primary paper per lecture
- Provider
- UC San Diego (Yiying Zhang)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How LLM inference splits into prefill and decode, and how continuous batching serves concurrent requests
- ✓How PagedAttention manages KV cache memory and how tool calls interact with KV state
- ✓GPU architecture basics and kernel programming ideas through the ThunderKittens framework
- ✓Running LLMs on edge devices with NPUs, heterogeneous execution and energy-aware adaptive serving
- ✓Systems support for agents: tooling, memory and context management, OS-level sandboxes and agent security
- ✓Distributed serving with prefix caching, prefill-decode disaggregation, long-context and MoE serving
- ✓Production infrastructure for serving agents, distributed training pipelines and RL fine-tuning systems like verl
Highlights
- •Every lecture has public PDF slides and a paper from a top systems venue (OSDI, SOSP, NSDI, MLSys, USENIX Security)
- •The syllabus is current to 2026, with NSDI '26, MLSys '26 and USENIX Security '26 papers on agent serving and security
- •It treats agents as a systems problem (sandboxes, cloning, serving), which framework-focused agent courses leave out
- •Taught by an OSDI Best Paper winner whose lab builds LLM serving and computer-use agent systems
Who It’s For
Best For
- ✓ML infrastructure and platform engineers who serve LLMs or run agent workloads in production
- ✓Systems engineers moving into LLM inference, GPU and serving work
- ✓Graduate students or researchers looking for a paper-by-paper reading list on LLM and agent systems
Prerequisites
- •Undergraduate-level operating systems (memory management, scheduling, virtualization)
- •Undergraduate-level machine learning, including how transformer models are trained and run
- •Comfort reading systems research papers; self-learners get slides and papers, and lecture recordings are not linked on the course page
FAQ
What is CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026)?
A Fall 2026 UC San Diego graduate course by Yiying Zhang covering the backend systems behind LLMs and AI agents. It suits engineers who already know operating systems and ML and want to understand inference, distributed serving, agent runtimes and training infrastructure. The public slides and paper list support self-study.
Is CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026) free?
CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026) is free to access.
What level is CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026) for?
CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026) is aimed at a advanced audience. Recommended background: Undergraduate-level operating systems (memory management, scheduling, virtualization), Undergraduate-level machine learning, including how transformer models are trained and run, Comfort reading systems research papers; self-learners get slides and papers, and lecture recordings are not linked on the course page.
How long does CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026) take?
Expect roughly 10 weeks, 17 lectures of 80 minutes; self-study via public slides plus ~1 primary paper per lecture. Most learners work through it at their own pace.
What will I learn from CSE 291A: Systems for LLMs and AI Agents (UC San Diego, Fall 2026)?
You'll learn: How LLM inference splits into prefill and decode, and how continuous batching serves concurrent requests; How PagedAttention manages KV cache memory and how tool calls interact with KV state; GPU architecture basics and kernel programming ideas through the ThunderKittens framework; Running LLMs on edge devices with NPUs, heterogeneous execution and energy-aware adaptive serving; Systems support for agents: tooling, memory and context management, OS-level sandboxes and agent security; Distributed serving with prefix caching, prefill-decode disaggregation, long-context and MoE serving; Production infrastructure for serving agents, distributed training pipelines and RL fine-tuning systems like verl.
Topics
Sources
This page was written from 4 sources, 1 on domains other than cseweb.ucsd.edu.