Design Docs Are All You Need: An AI-native Machine-Learning Performance Tool

Researchers have introduced a new machine learning performance modeling tool called SMART, aimed at decoupling code from documentation. The core mechanism of this tool is to construct design documents as directed acyclic graphs, and programming sub-proxies automatically generate code based on the documents, ensuring that humans can complete system iteration by simply editing natural language descriptions. SMART relies on two key elements to ensure reliability: first, a design document style based on step-by-step examples to provide generative context; second, a minimalist intermediate representation that includes SymPy cost expressions and recursive operator definitions. Tests show that this tool can highly accurately reproduce manually audited reference models (such as DeepSeek-V3 on TPU nodes), with accuracy reaching round-off level, confirming that design documents can serve as a durable engineering artifact for collaborative development of machine learning systems.