Introduction to LLM
This page provides an easy-to-understand guide on LLMs (Large Language Models) from basics to applications for AI enthusiasts.
Chapter 10 — The LLM Engine Layer
Tenth post of the LLM Primer VI walkthrough. vLLM as the safe default, TensorRT-LLM for peak NVIDIA-only throughput, SGLang for structured and agentic outputs, and TGI/Ollama for the rest.
2026-05-02
Chapter 7 — Advanced Batching Strategies
Seventh post of the LLM Primer VI walkthrough. Static vs dynamic vs continuous (in-flight) batching, iteration-level scheduling, and how a batch's slots actually progress on the GPU.
2026-04-29
Chapter 2 — The KV Cache Challenge
Second post of the LLM Primer VI walkthrough. The KV cache formula, the attention-variant trade-offs (MHA vs GQA vs MQA), and the memory-fragmentation problem PagedAttention solves.
2026-04-24
Chapter 1 — The Mechanics of Token Generation
First post of the LLM Primer VI walkthrough. The autoregressive bottleneck, the prefill/decode split, and why a high-end GPU is 99.7% idle while serving a single user.
2026-04-23
Chapter 4 — AI Agents and Tool Calling
Fourth post of the LLM Primer V walkthrough. ReAct loops, tool schemas as contracts, and the three memory layers agents actually need in production.
2026-04-17
Chapter 14 — Benchmarking, Testing, and Performance
Fifteenth and final post of the LLM Primer IV walkthrough. The MCP-Universe Benchmark on real servers, the two systemic failure modes it exposed, the ten-times throughput gap between session-per-request and shared session pools, and the bridge to Volume V.
2026-04-12
Chapter 10 — Long-Horizon Task Memory
Tenth post of the LLM Primer IV walkthrough. Short-term memory through windows and ReAct scratchpads, long-term memory through episodic vectors and semantic stores, and the compaction techniques that keep an agent productive over hours and days.
2026-04-08
Chapter 7 — Efficiency and Transformer Variants
Seventh post of the LLM Primer II walkthrough. The computational complexity of attention, the GPU memory and throughput math that constrains real systems, FlashAttention derived from first principles, and the family of clever variants — multi-query, gated, low-rank — that keep big models running.
2026-03-09
Chapter 11 — Cutting-Edge Research: MoE, Reasoning Models, and the New Scaling Axis
Chapter 11 of the LLM Primer I series. The research frontiers that are now production reality — mixture-of-experts, retrieval-augmented memory, native multimodal tokenization, continual learning, and the inference-time scaling paradigm that produced today's reasoning models. The 2026 edition's biggest content addition.
2026-02-28
Chapter 9 — Performance, Scaling, and Costs: The Real Engineering Trade-offs
Chapter 9 of the LLM Primer I series. The operational realities of running LLMs at scale — model size vs capability, the latency–throughput trade-off, cost economics, quantization, and edge deployment. Why frontier-tier models are often the wrong choice even when you can afford them.
2026-02-26
Chapter 5 — Training Large Models: What Actually Goes Into a Frontier Model
Chapter 5 of the LLM Primer I series. How frontier LLMs are actually trained — the data pipeline, the loss function, the months of GPU time, and why "training" is now an industrial-scale engineering problem more than a research problem. Demystifies what those hundred-million-dollar training runs are paying for.
2026-02-22