Course Schedule

Note that the schedule below is tentative and subject to change. Paper assignments will be finalized by the start of the semester. Optional papers are provided as additional reference, and we will cover some of them in class. They are not required as a part of the paper reviews.

Date	Topic	Paper	Assignment
8/21/25	ML Systems Overview and Logistics
8/26/25	Deep Learning Overview
8/28/25	AI Lifecycle and MLOps	Ray: A Distributed Framework for Emerging AI Applications (Optional) Meet Michelangelo: Uber's Machine Learning Platform (Optional) Hidden Technical Debt in Machine Learning Systems
9/2/25	[Guest Lecture: Prof. Francisco Romero, Georgia Tech] ML in Practice: Visual Data Analysis	Zelda: Video Analytics using Vision-Language Models
9/4/25	No Class
9/9/25	ML Libraries and Frameworks	TensorFlow: A System for Large-Scale Machine Learning (Optional) PyTorch: An Imperative Style, High-Performance Deep Learning Library (Optional) TVM: An Automated End-to-End Optimizing Compiler for Deep Learning (Optional) MLIR: Scaling Compiler Infrastructure for Domain Specific Computation
9/11/25	Hardware Acceleration	In-Datacenter Performance Analysis of a Tensor Processing Unit (Optional) Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures	Lab 1 Released
9/16/25	Training: Data Pipeline	cedar: Optimized and Unified Machine Learning Input Data Pipelines (Optional) Understanding Data Storage and Ingestion for Large-Scale Deep Recommendation Model Training
9/18/25	Training: Data Pipeline	Modyn: Data-Centric Machine Learning Pipeline Orchestration (Optional) Lakehouse: A New Generation of Open Platforms that Unify Data Warehousing and Advanced Analytics
9/23/25	Training: Parallelization	Alpa: Automating Inter- and Intra-Operator Parallelism for Distributed Deep Learning (Optional) Efficient Large-scale Language Model Training on GPU Clusters using Megatron-LM	Project Proposal Due
9/25/25	Training: Parallelization	PipeDream: Generalized Pipeline Parallelism for DNN Training (Optional) GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism
9/30/25	Training: Memory Optimizations	ZeRO-Infinity: Breaking the GPU Memory Wall for Extreme Scale Deep Learning (Optional) ZeRO: Memory Optimizations Toward Training Trillion Parameter Models
10/2/25	Training: Checkpointing and Resiliency	ReCycle: Resilient Training of Large DNNs using Pipeline Adaptation (Optional) CheckFreq: Frequent, Fine-Grained DNN Checkpointing	Lab 1 Due
10/7/25	Training: Datacenter Design and Scheduling	MAST: Global Scheduling of ML Training across Geo-Distributed Datacenters at Hyperscale (Optional) Applied Machine Learning at Facebook: A Datacenter Infrastructure Perspective
10/9/25	Reading Day (No Class)
10/14/25	Midterm Project Check-in		Midpoint Project Report
10/16/25	Inference: Overview	INFaaS: Automated Model-less Inference Serving (Optional) Serving DNNs like Clockwork: Performance Predictability from the Bottom Up
10/21/25	Intro to LLMs	Training Compute-Optimal Large Language Models (Optional) Attention is All You Need
10/23/25	LLMs: Inference Overview	Orca: A Distributed Serving System for Transformer-Based Generative Models (Optional) Fast Inference from Transformers via Speculative Decoding	Lab 2 Released
10/28/25	LLMs: KV Caching and Attention Optimizations	Efficient Memory Management for Large Language Model Serving with PagedAttention (Optional) FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness (Optional) SGLang: Efficient Execution of Structured Language Model Programs
10/30/25	LLMs: KV Caching and Attention Optimizations	DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving (Optional) Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve (Optional) Splitwise: Efficient Generative LLM Inference Using Phase Splitting (Optional) Mooncake: Trading More Storage for Less Computation — A KVCache-centric Architecture for Serving LLM Chatbot
11/4/25	LLMs: Compound Systems, Agents, and Reasoning	(Required) The Shift from Models to Compound AI Systems (Required) Hermes: Algorithm-System Co-design for Efficient Retrieval-Augmented Generation At-Scale
11/6/25	LLMs: Compound Systems, Agents, and Reasoning	Generative Agents: Interactive Simulacra of Human Behavior (Optional) Learning to Reason with LLMs
11/11/25	LLMs: Large-Scale Training, RLHF, Fine-tuning, and Post-training	MegaScale: Scaling Large Language Model Training to More Than 10,000 GPUs (Optional) Efficient Streaming Language Models with Attention Sinks (Optional) Scaling Llama 3 Training with Efficient Parallelism Strategies (Optional) DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning	Lab 2 Due
11/13/25	LLMs: Variants (MoEs, LoRAs, etc.)	S-LoRA: Serving Thousands of Concurrent LoRA Adapters (Optional) Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer
11/18/25	Security and Privacy	Privacy Side Channels in Machine Learning Systems
11/20/25	Responsibility and Sustainability	Sustainable AI: Environmental Implications, Challenges and Opportunities
11/25/25	Thanksgiving: No Class
11/27/25	Thanksgiving: No Class
12/2/25	Final Project Presentations
12/4/25	Final Project Presentations