NANOCHAT

Summary from the TLDR newsletter

nanochat is a minimal full-stack training/inference pipeline for large language models (LLMs) designed to run on a single 8XH100 node. It can train an LLM in as little as four hours. nanochat handles tokenization, pretraining, finetuning, evaluation, inference, and web serving.