comment stringclasses 1
value | framework stringclasses 1
value | contracts dict | environment dict | known_failures dict |
|---|---|---|---|---|
Pinned consolidated MIC snapshots in huggingface-mic/<name>-mic. Usage and framework expectations are in each immutable Hub snapshot. Stable fixture target names retain baseline identities. Run: run-model-contracts --fixtures-only --lock tests/contracts/fleet-lock.json | transformers | {
"SmolLM": {
"integration": {
"repo": "huggingface-mic/smollm-mic",
"revision": "39a8a2e580277c904e7d9737302b408d168e8efb",
"files_sha256": "fe4674584d161b963cc40d293e9086851f1e33ced8ef1de5d37005d57b5a402a"
},
"fixtures": {
"huggingface-mic/smollm-mic": {
"revision": "39a8... | {
"python": "3.10",
"torch": "2.11.0",
"torchvision": "0.26.0",
"torchaudio": "2.11.0",
"tokenizers": "0.23.1",
"numpy": "1.26.4",
"pillow": "12.2.0",
"accelerate": "1.13.0",
"safetensors": "0.8.0",
"huggingface-hub": "1.32.0",
"sentencepiece": "0.2.1",
"protobuf": "6.33.6",
"librosa": "0.11.0... | {
"Wav2Vec2/mic-wav2vec2-base-960h-slim": {
"issue": "https://github.com/huggingface/transformers/issues/49385: masked_spec_embed uninitialized, so the SpecAugment loss and gradients are undefined (fix #49386)",
"skip_outputs": [
"training.gradients",
"training.loss"
]
}
} |
Transformers MIC fleet
This dataset defines the Model Integration Contracts run by Transformers CI.
fleet-lock.json: the control fleet, run with the default variant.prototype-fleet.lock: the prototype fleet, run with the bfloat16 variant.
Each entry pins a huggingface-mic/<model>-mic model snapshot containing the slim
weights, executable usage and reviewed Transformer expectations. Contract and
fixture hashes verify the pinned content. Known failures are scoped in the locks.
The runner lives in transformers-ci.
Transformers pins the dataset commit SHA in tests/contracts.lock; CI uses
that revision for both fleet invocations. tmic export --execute updates this
dataset; tmic pin --transformers-checkout PATH --revision SHA adopts a revision.
Results and traces record that revision for reproduction. Model repository
revisions in the locks remain immutable even when this dataset is updated.
run-model-contracts --fixtures-only --fleet control
run-model-contracts --fixtures-only --fleet prototypes --variant bfloat16
# Reproduce a job with its recorded dataset commit:
run-model-contracts --fixtures-only --fleet prototypes --variant bfloat16 --fleet-revision <dataset-commit>
Update the locks after measuring and reviewing changed expectations. CI checks against expectations; it does not record or approve new baselines.
- Downloads last month
- 73