Opt-in CUDA, CPU default

[emrefbulut]

Summary

• CUDA is opt-in via --device cpu|cuda on iqforge train and iqforge measure-leakage. The default stays CPU even when a GPU is present; cuda errors if torch reports none.
• TrainingResult.environment (and measurement runs) record cpu vs cuda so two tables can be told apart. No GPU bit-equality is claimed.
• Published tables, scripts/parity_gate.py, and artifacts/*.json stay on CPU. The parity path does not pass --device. Seed lists and n are unchanged.

Why

ROADMAP optional CUDA for new measurements only. A silent GPU default would break the reproducibility promise and mix devices inside paired experiments.

How it was verified

uv run pytest tests/test_models.py tests/test_measurement.py tests/test_preflight.py — 53 passed, 2 skipped.

Mutation check: resolve_device was temporarily changed to pick CUDA whenever torch.cuda.is_available(). test_default_stays_cpu_even_when_cuda_is_available, test_cpu_flag_must_not_silently_use_cuda, and test_current_environment_stamps_an_opt_in_cuda_device went red, then the mutation was reverted.

Conventions it touches

• 1 Never fall back silently — --device cuda errors if torch reports no CUDA; a present GPU does not change the default
• 2 A passing test does not prove it can fail — CPU-default tests were turned red by the mutation above
• 3 Do not claim what you did not measure — environment stamps the device; CUDA runs warn they are not bit-comparable
• 4 Protect published artifacts — no file under artifacts/ was rewritten; the parity gate does not opt in to CUDA

Made with Cursor