One-Click LLM Training: A Distributed Computing Platform Based on HGX

Pavel, a senior ML platform developer at Avito, discusses the evolution of the company's infrastructure from individual SSH machines to the robust cloud-native Aviflow platform. The article details the transition to distributed LLM training, driven by the growth of the data science team and the need for optimized compute resources. The author shares experiences in implementing an HGX-based infrastructure, discusses the technical challenges of adopting MLOps practices, and explains how automating model training processes significantly improved specialist efficiency. This article is useful for engineers building ML platforms and those facing challenges in scaling compute for large language models. The material is based on a presentation from Kuber Conf and covers practical aspects of building modern ML infrastructure within a large enterprise.