Landscape Memo
Purpose
This memo documents why pytest-gpu-proof was built rather than adopting an existing tool.
Search terms used
pytest gpu test pluginpytest signed receipt attestationpytest artifact signingpytest equivalence testingGPU test attestation pythonpytest-json-report,pytest-artifacts(direct name searches)sigstore python signing pytestNVIDIA attestation SDK python- PyPI full-text search:
gpu proof,gpu attestation,gpu equivalence
Existing tools evaluated
pytest-json-report
What it does: Emits a JSON report of the test session (outcomes, durations, log output).
Gap: No signing, no code fingerprinting, no GPU-specific helpers. Useful as a complement to this plugin but does not replace it.
pytest-artifacts
What it does: Collects and uploads test artifacts (logs, screenshots) to a configurable store.
Gap: No signing, no equivalence checking, no GPU concept.
GitHub artifact attestations + Sigstore
What it does: Signs GitHub Actions workflow artifacts with Sigstore keyless signing, tied to the GitHub Actions OIDC identity and logged in the Rekor transparency log.
Gap: Works only inside GitHub Actions. Cannot sign a receipt produced on a local developer machine or lab GPU. Has no pytest integration layer, no equivalence testing helpers, and no verifier that understands pytest test outcomes.
Sigstore Python client (sigstore PyPI package)
What it does: Programmatic access to Sigstore signing and verification in Python.
Gap: A library, not a pytest plugin. No GPU equivalence concept, no receipt format, no pytest hooks. Suitable as a future optional signing backend for this plugin.
NVIDIA Attestation SDK (nv-attestation-sdk)
What it does: Hardware-level attestation for NVIDIA Hopper GPUs running in Confidential Computing environments. Produces cryptographic evidence that specific code ran on specific verified GPU hardware.
Gap: Requires Hopper-generation hardware and a Confidential Computing environment. Far too heavyweight for ordinary GPU correctness testing. Solves a different problem (hardware trust) than this plugin (team workflow trust).
pytest-randomly, pytest-benchmark, pytest-cov
None of these are relevant. Included for completeness; all solve adjacent problems.
Conclusion
No maintained package combines:
- GPU equivalence test helpers (reference vs. candidate comparison)
- Signed receipt / attestation output
- Local-first verification flow (SSH key, no cloud dependency)
- Optional CI-GPU execution mode
- GitHub-friendly verification policy (public key from
github.com/{user}.keys)
Building pytest-gpu-proof as a focused, narrow package is justified.
The implementation reuses cryptography (Ed25519, SSH key parsing) and standard
pytest hooks rather than inventing new infrastructure.