Topic 21: LLM Engineering
Phase: 5 | Status: ✅ Complete | Prerequisites: 13 Neural Networks, 16 Transformer
Overview
The modern LLM pipeline consists of pre-training, parameter-efficient fine-tuning (PEFT), and alignment. This module covers Byte-Pair Encoding (BPE) for subword tokenization, Low-Rank Adaptation (LoRA) and QLoRA for efficient fine-tuning, and Reinforcement Learning from Human Feedback (RLHF) along with Direct Preference Optimization (DPO) for model alignment.
Scope
- In scope: word-level BPE (the classic Sennrich et al. algorithm, matching
theory.mdand the tested package implementation), LoRA rank decomposition and parameter accounting, and the DPO objective in stable log-space — all in pure NumPy. Byte-level BPE is described as a real-world variant. - Out of scope: training or serving actual language models, QLoRA quantization internals, and RLHF reward-model training.
Contents
| # | File | Type | Description |
|---|---|---|---|
| 1 | theory.md |
Theory | BPE tokenization algorithm, LoRA derivations and cost analysis, RLHF pipeline, DPO objective derivation, failure modes |
| 2 | first_principles.ipynb |
Computation | WHY→BUILD→VERIFY — from-scratch NumPy implementations of BPE Tokenizer, LoRA Linear layer, and DPO loss |
| 3 | exercises.ipynb |
Practice | Hand calculation of BPE merges, LoRA parameter savings computation, conceptual analysis of DPO vs RLHF |
The reusable, unit-tested reference implementations live in src/ml_first_principles/llm_models.py (BPETokenizer with fit/encode/decode, LoRALinear, dpo_loss), covered by tests/test_phase5_models.py.
Connections
- Prereqs: 13 Neural Networks, 16 Transformer
- Synthesis: Connections to Optimization Methods Compared
- Next: Advanced inference optimization (KV caching, speculative decoding, continuous batching)