1 Ordinary tests

• Use an isolated checkout on a Linux GPU host, such as tom-workstation, in an official PyTorch CUDA devel container with an idle GPU exposed. Install pytest, nvidia-ml-py, and setuptools in the same Python environment as PyTorch.
• From the repository root, rebuild the extensions for the container's CUDA major, then run all ordinary tests. This example uses CUDA 12; set TMS_CUDA_MAJOR=13 for CUDA 13:

TMS_CUDA_MAJOR=12 uv run --no-project python setup.py build_ext --inplace
CUDA_VISIBLE_DEVICES=0 timeout 3600 uv run --no-project python -m pytest test -vv -ra

• The suite parametrizes supported scenarios over preload and torch hook modes. Single-GPU runs skip multi-device cases; report skips with the passing summary.
• If preload children cannot find libcudart.so.<major>, prepend the matching runtime directory to LD_LIBRARY_PATH before pytest; pip CUDA runtimes may be outside the loader's default search path.
• For build/release tooling changes, also run:

uv run --no-project --with pytest --with packaging --with typer --with setuptools python -m pytest \
  test/test_merge_cuda_wheels.py \
  .claude/skills/tms-publish-release/scripts/test_*.py -vv -ra

2 Full matrix tests

• Run the ordinary and tooling tests above, then use tms-publish-release Sections 2–5 for isolated paths, Docker/ARM64 binfmt setup, both wheel builds, and fresh GPU containers on tom-workstation.
• For development testing, use the target commit and its setup.py version; release-only requirements to use merged origin/master, check PyPI, or publish do not apply.
• After building both wheels into dist/, run on the GPU host with the paths prepared above:

uv run --script .claude/skills/tms-publish-release/scripts/gpu_validation.py \
  --release-root "$TMS_REMOTE_ROOT" \
  --artifact-root "$TMS_REMOTE_ARTIFACTS" \
  --expected-version "$TMS_RELEASE_VERSION"

• Require all four cells to pass: x86_64/ARM64 × CUDA 12/13. The harness tests installed wheels outside the checkout and enforces exact skips; do not deselect failing cases.
• When adding test/test_*.py, update the harness runtime/build-tool inventory and run every classified build-tool module; unclassified modules fail preflight.
• Multi-device, XPU, and Lupine exclusions remain coverage limits. Validate affected backends on suitable hardware separately.
• For either chapter, save complete build/test output in unique artifact logs, run long commands in the background, and inspect progress at least every ten minutes.