Volume 5

Volume 5 - Large Language Models

Volume 5 follows one model family, Meta's Llama 3.1, from the bytes of a prompt to a priced production fleet. Its first section reads the published configuration into tokens, embedding rows, rotary positions and attention heads; the second follows pretraining, fine-tuning, preference optimization and the alignment stack; the third works out how the KV cache, number formats, distillation and serving schedulers fit it onto H100, A100, L4 and RTX 4090 GPUs. Real configurations, papers and datasheets supply the numbers, and each chapter commits to a choice a team would have to make and says what would reverse it.