Optional extras version matrix (W14.T06)¶
Core lab (bnn-lab) does not require transformers or torchao. Those
extras are bridges / pedagogy only — never part of the 32× pack thesis gate.
| Extra | Pin (pyproject) | Probe | CI default |
|---|---|---|---|
[hf] → transformers |
>=4.40,<5 |
python scripts/smoke_optional_extras.py |
skipped unless [hf] installed |
torchao |
unpinned (GPU host recipe) | same script --probe torchao |
never required on CPU CI |
datasets / huggingface_hub |
with [hf] |
import smoke | optional |
How to run¶
pip install -e ".[hf]" -c constraints.txt # optional
python scripts/smoke_optional_extras.py
python scripts/smoke_optional_extras.py --require hf # fail if missing
Pytest marker:
pytest -q -m optional_extras tests/test_optional_extras_matrix.py
Honesty¶
- torchao INT4/FP8 is the commodity GPU path — not classic BNN XNOR.
See
24_GPU_INT4_FP8_LANE.mdandscripts/bridges/torchao_int4_recipe.py. - A green optional probe does not refresh
golden_floors.json. - Version skew between local HF and CI is expected; pins keep the band honest.
Status¶
Smoke script + skip-if-missing pytest shipped. Full multi-version matrix in CI
remains optional (workflow_dispatch) — CPU runners should not pull CUDA
torchao wheels by default.