Hours of YouTube and 85 handtyped pages of notes on LLM Inference – Start here

A GitHub repository provides a comprehensive educational resource on large language model (LLM) inference, featuring 6+ hours of video content and 80+ pages of handwritten notes that explain how to build a small inference engine in C++ and CUDA from first principles. The project walks through each component of transformer inference—from loading model weights to generating tokens—and includes a working implementation for running Llama 3.2 1B with features like paged attention, KV caching, and batching optimization.