GPU validation
The CUDA checks execute PyTorch and CuPy in separate child interpreters. A native abort, timeout, import failure or nonzero worker exit produces a failed record while the parent interpreter remains alive.
For the package installed in the selected environment:
python -m komorebi_math verify gpu
Checked computations
The PyTorch worker checks CUDA availability and evaluates a cubic directional
JVP at 2, with value 8 and tangent 12. It also checks a device matrix-vector
product with expected result [14, 38, 62, 86].
The CuPy worker checks device availability, executes a device-resident
PacketIR, and compares the result with [0, 1, 4, 6] using Euclidean residual
norm at most 1e-12.
These are bounded equivalence controls for the recorded environment. They do not establish unrestricted GPU performance, stress coverage or compatibility with every driver and framework build.
Release records
The hardware producer verifies the installed wheel inventory before and after execution, clears source-path overrides, records interpreter/device/runtime versions and retains the nested worker results. Its temporary installation uses the exact candidate wheel and inherits the selected host's CUDA runtimes.
Follow the release guide for that artifact-bound command. Keep the complete report, including failed runs. An installed-package diagnostic and a release-bound record answer different questions.