LLM Avalanche: Xingyang Gang : Building Simple and Scalable Training Framework for LLMs

ai.bythebay.io Nov 2025, Oakland, full-stack AI conference In this talk, we will cover our recent effort in building EasyLM, a simple and scalable open source framework for training large language models. We discuss the benefits of building simple and easy to use LLM frameworks, and how we can leverage ML compilers to scale it up. We present two success stories of training LLMs with EasyLM: Koala, the dialogue model for academic research and OpenLLaMA, our permissively licensed reproduction of LLaMA.