A checkpoint resume from post_imputation successfully materialized PE variables and saved a post_microsim checkpoint, then failed when the microcalibrate backend attempted to use torch tensor operations in an environment where torch had been stubbed only to unblock non-neural imports.
Resume command shape:
.venv/bin/python -m microplex_us.pipelines.pe_us_recalibrate_from_checkpoint \
--checkpoint-path artifacts/local_us_microplex_smoke/local-smoke-v1/checkpoints/post_imputation \
--output-root artifacts/local_us_microplex_smoke/local-smoke-v1/recalibrated_from_post_imputation \
--targets-db /Users/administrator/Documents/PolicyEngine/policyengine-us-data/policyengine_us_data/storage/calibration/policy_data.db \
--target-period 2024 \
--calibration-backend microcalibrate \
--policyengine-materialize-batch-size 100000 \
--pipeline-checkpoint-save-post-microsim-path artifacts/local_us_microplex_smoke/local-smoke-v1/checkpoints/post_microsim
Progress before failure:
US microplex build: post-microsim checkpoint saved [path=artifacts/local_us_microplex_smoke/local-smoke-v1/checkpoints/post_microsim]
Failure:
File ".venv/lib/python3.14/site-packages/microcalibrate/calibration.py", line 117, in __init__
torch.manual_seed(self.seed)
File ".venv/lib/python3.14/site-packages/torch/__init__.py", line 58, in _unavailable
raise RuntimeError(...)
RuntimeError: The local smoke environment is using a torch stub. Install a real torch wheel before using neural synthesis or tensor operations.
The torch stub was used only because the project selected Python 3.14 and uv sync --extra policyengine could not install a real torch wheel on this workstation. The canonical smoke command uses --calibration-backend microcalibrate, so the environment still needs a real torch install even though neural synthesis is not used.
Next retry can resume from artifacts/local_us_microplex_smoke/local-smoke-v1/checkpoints/post_microsim, skipping donor integration and PE materialization.
A checkpoint resume from
post_imputationsuccessfully materialized PE variables and saved apost_microsimcheckpoint, then failed when themicrocalibratebackend attempted to use torch tensor operations in an environment where torch had been stubbed only to unblock non-neural imports.Resume command shape:
Progress before failure:
Failure:
The torch stub was used only because the project selected Python 3.14 and
uv sync --extra policyenginecould not install a real torch wheel on this workstation. The canonical smoke command uses--calibration-backend microcalibrate, so the environment still needs a real torch install even though neural synthesis is not used.Next retry can resume from
artifacts/local_us_microplex_smoke/local-smoke-v1/checkpoints/post_microsim, skipping donor integration and PE materialization.