Truly Serverless GPUs: A Deep Dive Inside Modal's Fast Cold Starts

Fast cold starts make the difference between an economical service with happy users and a world of pain. In this webinar, Member of Technical Staff Charles Frye will present the core technologies (custom container runtime + filesystem, CPU checkpoint/restore, and GPU checkpoint/restore) that allow applications on Modal to achieve >40x faster cold starts, e.g. starting an SGLang server in 10 seconds or less.