Checkout f9ed2bdaede7b66e6323dc7772a93953af39dfc7, build its cuda_core from source in a dedicated environment. Tested with cuda-bindings 13.2.0, CUDA Toolkit 13.2.78, libNVVM 2.0, IR 2.0.3.2, aarch64/SM110.
LD_LIBRARY_PATH=/usr/local/cuda-13.2/lib64:/usr/local/cuda-13.2/nvvm/lib64 \
python b3_compile_audit.py --root /path/to/cuda-python --out /new/output-directoryThe script imports the frozen checkout's existing NVVM textual/bitcode fixtures. It independently lists all 55 public fields and checks them against the dataclass inventory. It compiles each of the twelve reported fields with None/explicit values, including boolean False/True, through no-cache/cache-miss/cache-hit paths; the wrapper counts calls to the real uncached compiler. It also probes six flags directly through libNVVM for both IR formats. PCH/trace side effects are confined to a temporary directory.
Observed: 182 public API rows compiled/returned cached PTX without warnings. Sixty uncached rows and sixty misses called the real compiler once each; sixty hits called it zero times. Two baseline rows also compiled. Direct Ofast 0/min/mid/max all compiled for both formats; direct device-float128 and pch were rejected as invalid options. See attached raw JSON.
The field classification is a current-behavior audit, not an approved new API contract. Compiler support outside this version/fixture is not established; no wrong-cache-hit is claimed. Warning versus rejection for previously silent NVRTC-only settings awaits maintainer agreement, and Ofast needs separate treatment. No public behavior patch has been applied.