Design Docs Are All You Need: An AI-native Machine-Learning Performance Tool
SMART treats natural-language design docs as the durable source artifact, with code regenerated from them on each version update.
The paper describes a symbolic ML performance-modeling library whose main branch contains almost no code. Its implementation is rebuilt by coding sub-agents from a DAG of self-contained design docs, while human changes are made as edits to those docs. The authors say reliability comes from worked examples in the docs and a compact operator IR with SymPy cost expressions, analytical roll-up, and modulo-scheduling modes. Regenerated models matched hand-audited references, including DeepSeek-V3 serving on a TPU pod slice, to round-off precision. ArXiv · AI/CL/LG's note
The paper describes a symbolic ML performance-modeling library whose main branch contains almost no code. Its implementation is rebuilt by coding sub-agents from a DAG of self-contained design docs, while human changes are made as edits to those docs. The authors say reliability comes from worked examples in the docs and a compact operator IR with SymPy cost expressions, analytical roll-up, and modulo-scheduling modes. Regenerated models matched hand-audited references, including DeepSeek-V3 serving on a TPU pod slice, to round-off precision. ArXiv · AI/CL/LG's note
score 4