Hugging Face proposes Model Integration Contracts for Transformers, using shrunk models to catch regressions on CPU

Hugging Face proposed transformers-mic, packaging model tests as Model Integration Contracts (MICs): a shrunk model, a runnable usage example and human-reviewed baseline outputs, run on CPU. Replaying five months of commits, a 66 MB UMT5 fixture (source checkpoint 1,227 MB) flagged a bug that sat in main for almost two months: with SDPA enabled, UMT5's decoder could attend to future tokens.

Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.