Transformers Model Integration Contracts

Proposed Hugging Face standard that packages shrunk-model integration tests to run cheaply on CPU in Transformers CI.

Hugging Face · Evaluation, observability, and MLOps tooling · open-source

Published in Feed

Overview

Transformers Model Integration Contracts (MICs), described in a November 3, 2026 Hugging Face post by Tarek Ziadé, bundle a small fixture model built from a real checkpoint's configuration, a runnable usage path and reviewed baseline outputs; a contract holds if the usage path on the fixture still matches the baselines. The post cites a UMT5 bug that let the decoder see future tokens and went unnoticed from July 30 to September 28, 2026. Tooling includes the `tmic` CLI, the `transformers-mic` package, the `huggingface-mic` Hub org and a coverage Space. As of November 1, 2026 the MIC job in Transformers CI reports failures without blocking pull requests.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Three parts per contract: pinned fixture, usage path and baselines
  • Shrinks models while keeping layer kinds: Llama 4 Scout goes from 108.6B to 44.7M parameters; the UMT5 fixture is 66 MB versus 1,227 MB
  • Fixtures must use the default SDPA attention path, since eager-only runs missed the UMT5 bug
  • Whisper fixtures trained on synthetic tones caught 24 of 24 injected faults versus 18 of 24 with random weights
  • `tmic` CLI drafts contracts, builds fixtures and records baselines, with human review at each step

Use cases

  • Catching model-implementation regressions on a pull request without GPUs or full checkpoints
  • Adding regression coverage for a newly contributed model type in Transformers, which has 750+ model types

Pricing

Free and open source; the Transformers CI job is currently non-blocking.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
3 Nov 2026
Record updated
3 Nov 2026
Record last checked
3 Nov 2026

Explore the Feed

Other entries filed under evaluation, observability, and mlops tooling; inclusion does not imply equivalent features.

Browse the Feed