Skip to content

fix: enforce the CPU-only automatic qualification boundary - #21

Open
EmilienIT wants to merge 3 commits into
magiccodingman:mainfrom
EmilienIT:codex/fix-auto-cpu-boundary
Open

EmilienIT wants to merge 3 commits into
magiccodingman:mainfrom
EmilienIT:codex/fix-auto-cpu-boundary

Conversation

@EmilienIT

@EmilienIT EmilienIT commented Oct 18, 2026 •

Copy link
Copy Markdown

Synchronized on 2026-10-18: current head 097f8bc36a7f23792df187234532691ade74cc96 includes upstream/main 1d2e24849c9d8441b9614acade5252c9937a2bbf through a regular merge. Proposal fixes and regression files remain unchanged. All proposal tests and the new upstream client-tool CPU/live and grammar upgrade/release gates are retained. Local baseline PR checks and complete portable Python compilation passed. Earlier revision-specific results below remain historical evidence.

Fresh complete Ubuntu 24.04/Python 3.12 CPU integration passed: https://github.com/EmilienIT/vllm-radiance/actions/runs/39142861907, integration head e840f823c973e14c17e990fec52b0d2deb65bbae. Repository 259 passed/19 existing native skips and 188 subtests passed; conformance 198 passed; D7 repair 230 passed/1 existing artifact skip; NVFP4 conversion 10 passed; FP8/TunableOp source checks passed. Total: 697 passed cases and 20 explicit existing skips, plus baseline/artifact checks and subtests. This is combined-proposal CPU evidence, not a standalone upstream run.

The contributor has no R9700. Candidate-image, native ROCm/HIP, physical TP2/TP4, performance and release qualification remain NOT RUN. Maintainer hardware/runtime qualification is required before promoting runtime changes or publishing an image.

run_auto currently uses Python truthiness for registry eligibility and never verifies the hardware boundary. A quoted YAML automatic: "false" is truthy, and an entry mistakenly marked automatic: true for r9700 or maintenance-window-r9700-tp2 is executed. The validator also accepts these configurations. This contradicts the existing automatic-CPU/manual-hardware contract.

Require boolean eligibility flags, reject non-CPU automatic entries, and preflight the complete selected plan before launching its first command. Enforce the same invariants in control-plane validation. Valid CPU automatic entries still execute, manual hardware entries remain explicit, and dry runs execute nothing. All currently registered checks already satisfy these constraints.

Proposal head: 8da55d3a69748f72d245ff1f498123dcbd8b3a86, based on upstream 1d9016bbcd5136b12a11dee285ffb614d9b577e3.

Validation:

  • Eight CPU tests exercise the actual runner and validator, with process launches mocked and asserted. Original source: 27 assertion failures including subcases; fixed source: all pass. The tests cover string/integer/null flags, every non-CPU hardware class, whole-plan preflight, manual entries and dry runs. No hardware probe was launched.
  • Windows/Python3.13: baseline and changed-area automatic checks passed. Complete source compilation is covered in the integration by the portable checker from fix: make Python syntax qualification portable across host shells #17, compensating for upstream's native-Windows glob warning.
  • Full Ubuntu24.04/Python3.12 CPU integration run passed, integration head e3f4437ded0c3bc633971f62204605d2945ac2d1: repository tests 167 passed / 19 existing native skips; conformance 198 passed; D7 CPU repair 230 passed / 1 existing compiler-artifact skip; NVFP4 CPU conversion 10 passed; FP8/TunableOp artifact source checks passed. git diff --check passed.
  • The changed run_auto function is AST-identical to the integration; validator, regression tests and policy text match byte for byte. The integration additionally contains fix: include local Codex edits in test selection and add CPU PR checks #16's pending selector/workflow and the independent review proposals; this is not a standalone upstream run. All regression IDs are retained in its registry.
  • The full conformance suite failed collection on native Windows (fcntl unavailable); the failure is retained. It passed in full on Ubuntu, without excluding those tests.

The contributor has no R9700. Candidate-image, ROCm/GPU, dual-R9700 TP2, live correctness, performance and release qualification: NOT RUN. CPU-qualified maintenance behavior only; no hardware qualification is implied. Maintainer physical-GPU qualification remains required before publishing runtime changes. This change neither modifies runtime defaults nor weakens a manual hardware gate.

Includes the mechanical #16 rename of the inherited comment-only Docker workflow to .yml.disabled, preventing invalid-workflow failure notifications. No executable check is disabled, and GitLab publication is unchanged. Original failed runs are preserved. No merge or image publication performed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant