• Sources: primary
  • Summary: Enabling SDPA for T5 on 2026-07-30 let UMT5's decoder attend to future tokens, and the bug stayed in main until 2026-09-28. A Model Integration Contract pairs a shrunk fixture model, a runnable usage path, and human-reviewed baselines, and a 66 MB UMT5 fixture flagged the regression on every replayed commit while an eager-attention variant missed it. As of 2026-11-01 the MIC job reports failures in Transformers CI without blocking PRs, and LLM-drafted contracts need human approval.
  • Why it matters: Teams testing ML code can catch silent numerical regressions on cheap CPU runners if their fixtures exercise the same default code paths as production checkpoints.