| """Shared safety scaffolding for the opt-in fused-kernel installers (CCE, custom RMSNorm). | |
| Both run a numeric parity check on the live GPU before patching and only patch if it matches eager | |
| within tolerance — correctness over speed. The check must not perturb the trainer's global RNG (it | |
| runs after the model is loaded, inside the trainer's RNG stream), so it seeds under a local fork. | |
| """ | |
| from __future__ import annotations | |
| from collections.abc import Callable | |
| def run_gpu_self_test(body: Callable[[], bool], *, seed: int = 0) -> bool: | |
| """Run ``body`` (a numeric-parity check returning bool) on the live GPU under a forked RNG. | |
| Returns False when CUDA is unavailable. Seeds with ``seed`` inside ``torch.random.fork_rng`` so | |
| seeding the check doesn't disturb the trainer's global CPU/CUDA RNG stream. | |
| """ | |
| import torch | |
| if not torch.cuda.is_available(): | |
| return False | |
| with torch.random.fork_rng(devices=[torch.cuda.current_device()]): | |
| torch.manual_seed(seed) | |
| return body() | |