mars

HPC Claim Review Checklist

Use this checklist before opening any external HPC-related PR, package submission, or public release note.

Disallowed Terms by Contract Level

Before any submission or publish-ready wording is approved, confirm the following term restrictions:

• GPU, CUDA, ROCm, Metal, or TPU claims require the project to be at least H3.
• accelerator claims require H3 unless explicitly framed as H3+-future work.
• MPI, multi-node, or multi worker claims require H4.
• distributed claims require H4 unless explicitly framed as a H4 non-goal.

This policy is enforced by scripts/check_hpc_claims.py through regex rules and can be reviewed with:

uv run python3 scripts/check_hpc_claims.py --strict

Reviewer Checklist by Contract Level

H0 - HPC-Packaging Ready

• Confirm submissions only claim source-installability, smoke-testability, and package-availability requirements.
• Confirm claims do not mention accelerator, distributed, or GPU execution in H0-scoped text.
• Confirm package identity and release versions are concrete (no placeholders).
• Confirm external links and track notes do not imply H1+ claims.

H1 - CPU Throughput Runtime

• Confirm deterministic single-thread fallback exists.
• Confirm thread-count controls are resource bounded and documented.
• Confirm benchmark baselines include single-thread and multi-thread medians plus a clear threshold policy.
• Confirm parallel paths preserve serial behavior and error semantics.
• Confirm serial-vs-parallel fixture parity tests are present.

H2 - Stable Runtime Boundary

• Confirm runtime-boundary documentation distinguishes API surfaces and stability expectations.
• Confirm memory ownership, error signaling, and versioning contract are explicit.
• Confirm Arrow/ABI interoperability claims are supported by docs and tests.
• Confirm at least one non-Python binding path has boundary-facing tests.

H3 - Accelerator-Ready Runtime

• Confirm the H3 accelerated backend is optional, not mandatory.
• Confirm CPU fallback remains functional and default-capable.
• Confirm H3 capability checks and unsupported feature behavior are documented.
• Confirm numerical tolerances are explicit for backend parity in the H3 scope.
• Confirm no H3 accelerated package creates hard installation requirements.
• Confirm H3 benchmark wording distinguishes the array-test validation backend from vendor H3 GPU/TPU/FPGA/ASIC speedup claims.
• Confirm H3 CUDA, ROCm, Metal, TPU, FPGA, and ASIC wording still names optional adapters unless vendor-specific parity and benchmark evidence is present.

H4 - Distributed Execution

• Confirm partitioning, ordering, and deterministic aggregation are specified.
• Confirm failure and retries semantics are documented.
• Confirm H4 cluster smoke/aggregation evidence is present.
• Confirm no implicit cluster start-up in default import/runtime behavior for H4.
• Confirm H4 command-backed multi-node claims mention explicit worker-command configuration rather than implicit cluster provisioning.

Reviewer Sign-off

For an external-ready submission, record:

• supported contract levels,
• exact evidence files referenced,
• any blocked claims explicitly preserved as non-goals or deferred work.

This site is open source. Improve this page.