CME295 Lecture Notes

이 폴더는 CME295 강의 영상을 기반으로 정리한 한국어 lecture notes 모음이다. 각 강의 README는 강의 요약, 핵심 개념, 실무 관점의 메모, 복습 질문과 답변, 그리고 Mermaid/SVG 다이어그램을 포함한다.

Lectures

Lecture	Topic	Notes	Source
01	Transformer 기초	lec-01/README.md	YouTube
02	Transformer-based models and tricks	lec-02/README.md	YouTube
03	LLMs, decoding, prompting, and inference	lec-03/README.md	YouTube
04	LLM training, fine-tuning, and efficient adaptation	lec-04/README.md	YouTube
05	LLM tuning and human preferences	lec-05/README.md	YouTube
06	LLM reasoning and GRPO	lec-06/README.md	YouTube

Learning Path

%%{init: {"theme": "base", "themeVariables": {"background": "#171717", "primaryColor": "#232323", "primaryTextColor": "#f5f5f5", "primaryBorderColor": "#d0d0d0", "lineColor": "#cfcfcf", "fontFamily": "Inter, Arial, sans-serif"}}}%%
flowchart LR
    A[Lecture 1<br/>Transformer basics] --> B[Lecture 2<br/>Transformer variants]
    B --> C[Lecture 3<br/>LLM inference]
    C --> D[Lecture 4<br/>LLM training]
    D --> I[Lecture 5<br/>Preference tuning]
    I --> J[Lecture 6<br/>LLM reasoning]
    A --> E[Attention<br/>Q/K/V]
    B --> F[Position, norm,<br/>BERT, KV cache]
    C --> G[MoE, decoding,<br/>prompting]
    D --> H[Pre-training, ZeRO,<br/>SFT, LoRA]
    I --> K[RLHF, PPO,<br/>DPO]
    J --> L[GRPO, pass@K,<br/>DeepSeek-R1]

    classDef primary fill:#232323,stroke:#d0d0d0,color:#f5f5f5,stroke-width:2px;
    classDef secondary fill:#3b2f20,stroke:#d0d0d0,color:#f5f5f5,stroke-width:2px;
    classDef note fill:#52676b,stroke:#d0d0d0,color:#f5f5f5,stroke-width:2px;
    classDef accent fill:#62164d,stroke:#d0d0d0,color:#f5f5f5,stroke-width:2px;
    class A,B,C,D,I,J primary
    class E,F,G,K,L note
    class H accent

Lecture Overview

Lecture 1: Transformer

Lecture 1은 NLP task, tokenization, representation learning, RNN/LSTM의 한계, attention의 필요성을 거쳐 Transformer encoder-decoder 구조를 설명한다. 핵심은 self-attention이 token 간 의존성을 병렬적으로 계산하고, query/key/value 구조로 information retrieval처럼 동작한다는 점이다.

주요 다이어그램:

• self-attention-qkv.svg
• transformer-encoder-decoder.svg

Lecture 2: Transformer-Based Models and Tricks

Lecture 2는 Transformer를 실제 model family로 확장할 때 필요한 positional encoding, RoPE, layer normalization, RMSNorm, MHA/MQA/GQA, encoder-only model, BERT pre-training을 다룬다. Lecture 1이 architecture의 기본 원리를 설명했다면, Lecture 2는 modern Transformer를 안정적이고 효율적으로 만드는 설계 선택을 정리한다.

주요 다이어그램:

• rope-rotation.svg
• mha-mqa-gqa-kv-cache.svg

Lecture 3: Large Language Models, Decoding, Prompting, and Inference

Lecture 3은 decoder-only Transformer가 LLM으로 확장되는 과정을 다룬다. Mixture of Experts, next-token decoding, greedy/beam/sampling, temperature, prompt structure, in-context learning, chain of thought, KV cache, PagedAttention, speculative decoding 등 inference-time behavior와 serving optimization이 중심이다.

주요 다이어그램:

• moe-routing.svg
• kv-cache-decoding.svg

Lecture 4: LLM Training, Fine-Tuning, and Efficient Adaptation

Lecture 4는 LLM을 어떻게 학습하고 조정하는지 설명한다. Pre-training, scaling laws, FLOPs/FLOP/s, GPU memory footprint, data parallelism, ZeRO, model parallelism, FlashAttention, mixed precision, SFT, instruction tuning, evaluation, alignment, LoRA, QLoRA가 핵심이다.

주요 다이어그램:

• zero-sharding.svg
• flashattention-io.svg
• mixed-precision-training.svg
• lora-qlora-adaptation.svg

Lecture 5: LLM Tuning and Human Preferences

Lecture 5는 SFT model을 human preference에 맞게 조정하는 preference tuning을 설명한다. Pairwise preference data, reward model, Bradley-Terry formulation, RLHF, PPO clip/KL penalty, reward hacking, best-of-N, DPO가 핵심이다.

주요 다이어그램:

• preference-tuning-pipeline.svg
• rlhf-dpo-tradeoff.svg

Lecture 6: LLM Reasoning and GRPO

Lecture 6은 reasoning model을 answer 전에 reasoning chain을 생성하는 LLM으로 보고, math/code처럼 정답 검증이 가능한 task에서 verifiable rewards와 GRPO로 reasoning behavior를 학습하는 방법을 설명한다. pass@K, sampling temperature, reasoning token cost, output length growth, DeepSeek-R1-Zero/R1 training pipeline, reasoning distillation이 핵심이다.

주요 다이어그램:

• reasoning-token-budget.svg
• grpo-group-advantage.svg

Concept Map

Concept	First covered	Later use
Tokenization	Lecture 1	pre-training data, SFT loss masking
Self-attention	Lecture 1	MHA, KV cache, FlashAttention
Q/K/V	Lecture 1	MHA/MQA/GQA, RoPE, KV cache
Positional information	Lecture 2	long context, RoPE, context rot
Transformer families	Lecture 2	decoder-only LLMs
Decoder-only LLM	Lecture 3	pre-training and SFT
MoE	Lecture 3	expert parallelism
KV cache	Lecture 3	inference memory and throughput
Scaling laws	Lecture 4	model/data/compute allocation
ZeRO	Lecture 4	distributed training memory
FlashAttention	Lecture 4	exact attention with lower HBM IO
LoRA/QLoRA	Lecture 4	efficient fine-tuning
Preference tuning	Lecture 5	human preference alignment after SFT
Reward model	Lecture 5	RLHF and best-of-N scoring
PPO	Lecture 5	RLHF policy optimization
DPO	Lecture 5	supervised-style preference optimization
Chain of thought	Lecture 3	reasoning chains and test-time compute
Preference tuning / PPO	Lecture 5	GRPO comparison and reasoning RL
pass@K	Lecture 6	coding/math reasoning evaluation
GRPO	Lecture 6	reasoning model RL training
DeepSeek-R1	Lecture 6	multi-stage reasoning training pipeline

Repository Notes

• Lecture notes are written in Korean, while technical terms are often kept in English when that is clearer.
• SVG diagrams use the shared editorial style from AGENTS.md: restrained palette, thin line boxes, minimal fill, and accent red only for critical paths.
• Mermaid diagrams include local classDef styling so they follow the same visual scheme in GitHub-rendered markdown.
• Each lecture README ends with review questions and answers for quick self-checking.