Repository navigation
Add cloud correctness gates and paired CPU scheduling benchmarks - #46
Conversation
|
Important Draft PR not reviewedDraft PRs are not automatically reviewed by default.
To automatically review draft PRs, update your CodeRabbit configuration: reviews:
auto_review:
drafts: true
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
The newly authorized integration pass admits this PR as tooling/evidence, with no production speedup claim. Runtime integration will merge PR46 then PR58, preserving Git ancestry. Both exact heads have green hosted feature/executable/scoped checks; PR46 also has green cloud acceptance. Combined local tree has no changes to src/, tests/, Cargo.toml or Cargo.lock. Formatting and cloud harness negative tests pass in normal and optimized Python; full local core regression and Clippy are still running as additional combined-tree verification. Book owner coordination is posted on PR43. No book branch or qualification gate changes in this slice. Main remains2209012 at preflight. Other runtime PRs remain held for the publication-source blocker; PR59 remains held for performance acceptance. This current user-authorized integration decision supersedes historical draft-only/no-merge instructions from the original implementation phase. |
Rheon needs repeatable CPU evidence that checks numerical acceptance before interpreting scheduling timings. This adds a clean-source qualification driver, a CPU step-plus-owned-capture example, adversarial harness tests, and a hosted workflow covering 16³/32³ grids and adaptive versus existing unbound batch sizing.
The driver retains raw logs, source/compiler/binary identity, CPU affinity/quota, three fresh-process repetitions per pair, per-child CPU/wall time and peak RSS, simulation/capture payloads, and complete-trajectory fingerprints. Existing solver/reference, conservation/refinement and shared dense3d contracts remain unchanged. Correctness and measurement validity are regression gates; timings are observational.
Validation:
cargo clippy --locked --features parallel --all-targets -- -D warningspassed.cargo test --locked --no-default-featurespassed, including the long material regression cases.1f79a4d(16³/32³; 1/2/4 workers; both pressure methods; fixed/adaptive; three repetitions; twelve steps). All trajectories advanced 0.24 seconds, with matching fingerprints across policies/threads and maximum divergence 1.01e-7. Peak process RSS was 14,720 KiB.Independent review findings were fixed: clean source, strict measurement parsing, actual parallel-size CI coverage, process-group timeout cleanup and runner-context placement. No tests were weakened.
See acceptance plan and measured evidence for exact commits, commands, raw-log locations, timing ranges and provenance. CPU timings exclude rendering/GPU/network transport; no real-device frame budget, general adaptive speedup, continuum validation or fresh Lean kernel qualification is claimed. Real AI evidence is limited to its documented native 32³ pressure-stage domain. All evidence is tied to its recorded source revision.
Continuation at
b395132adds independent consumer validation of the published API flow: hand-derived pressure, binomial donor transport, independently recomputed divergence and open-boundary volume/mass accounting. Actual nine-frame export passed, with zero fraction error. Three real-artifact/adversarial tests passed in normal/optimized modes. An external Rust recovery probe passed late cancellation, expired deadline, invalid input and partial export failure, then matched every remaining frame byte-for-byte with an uninterrupted trajectory. See continuing flow qualification.Exact-head cloud/executable/metadata checks passed at
60c15ad; new-head checks are tracked separately. Independently ran the four headless engineering/CAD tests at PR503f49a15, all passed; this does not qualify GPU/rendered performance. At combined PR53ec9d960, 19 selected release contracts passed with the real model enabled, all three session examples executed, and the independent trajectory oracle plus recovery probe passed. This qualifies the separate transport, material and pressure adapters; a production full-step AI hook remains unqualified.At AI PR49
4e4383c, the native 24-frame component trajectory, independent verifier and nine artifact tests in each Python mode passed. Native process cost was 206.233 s wall, 206.205 s CPU and 14,208 KiB peak RSS. However, independent rejection qualification failed: the verifier accepts zero timing receipts, unbounded reported timed residuals and duplicated frame evolution. The new black-box probe reproduces all four corruption cases, independently reviewed, without modifying original artifacts. This is a verifier defect, not evidence that original evolution is wrong. The historical failures and original evidence are preserved; see the fixed-verifier qualification below; measured rollout costs exclude field capture and cannot qualify complete exported-frame or game-engine performance.Current PR46 head
bdce233contains the regression probe and exact evidence. Hosted checks for this head are tracked separately from prior passing runs.No solver implementation changes, merge or deployment.
Fixed-verifier qualification: AI PR49
89997c5b9bb121c96f1fe40a7b050c130cc525b7leaves native sources, examples and Cargo inputs unchanged. The unchangedprobe_ai_trajectory_receipts.pypasses in normal and optimized Python against the original 291-file native capture: baseline accepted and all four mutations rejected. The owner's expanded trajectory suite passes 14 tests in each mode; pressure-stage suite passes seven in each mode with actual checkpoint/output. Independent source review found no blocking issue. A SHA256 manifest confirms every original trajectory artifact is unchanged (d10dabdd2a3df632f9c2e6d2981f24d893e847c4f053fa5005af87fd8ee0f6e8).Exact commands used
OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1with bothpython3andpython3 -O:tools/probe_ai_trajectory_receipts.py <fixed-verifier> /tmp/rheon-ai-trajectory-independent/trajectory;RHEON_AI_TRAJECTORY=<same-output> python3 -m unittest discover -s <fixed-tools> -p test_ai_pressure_trajectory.py -v; andRHEON_AI_CHECKPOINT=/tmp/rheon-ai-independent/model.pt RHEON_AI_RUN=/tmp/rheon-ai-independent/run python3 -m unittest discover -s <fixed-tools> -p test_ai_pressure_tools.py -v. Logs and probe JSON remain under/tmp/rheon-ai-{trajectory,pressure}-fixed-*.This qualifies consistency of captured evolution with the declared discrete scene, using pointwise engineering roundoff allowances. Timed repetitions retain receipt-consistency checks only; it does not qualify continuum physics, generalization, full exported-frame budgets or production-session recovery. Native timing evidence is reused because the verifier fix does not change executable source.
Terminal hosted status on exact PR46 head
bdce2338cc2bc623112af7fd8af35d2cc75655cc: all checks passed. Rust feature/executable workflow completed successfully: core-only 7m43s, default 8m20s, desktop 13m47s, executable 2m36s, aggregatecore-and-executablepassed. Cloud acceptance passed in 2m31s; both scoped metadata checks passed. The branch remains clean and pushed at that exact head; no further commit invalidates these checks.Production integration follow-up (validation commit
e245bcb): independently exercised PR55 exact3525c979e16629b295a6865e3abbb2a36803cc93. Twenty-nine selected release tests passed with the real model. New external recovery and field-transition probes accepted 24 heterogeneous-fraction intervals (23 learned, one explicit fallback); four rejection types preserved full publication/provider metadata and every retried frame matched an uninterrupted reference. Automatic timestep shortening reached the model support gate exactly. Both Python modes accept real captures and reject five corruption types. See production evidence and exact commands.PR55 native AI, executable, rational and metadata checks pass; feature jobs are being tracked to terminal. The book build fails; local reproduction confirms the pre-existing historical Cargo.lock source-pin refusal, with Cargo.lock unchanged from starting main. No production runtime defect found or gate weakened. This commit adds only validation tools/evidence and leaves PR55 source unchanged. Earlier all-green PR46 results above apply to
bdce233; they are not substituted for this new validation head.Terminal PR55 status on exact
3525c979e16629b295a6865e3abbb2a36803cc93: feature/executable workflow passed (core-only 14m7s, default 11m15s, desktop 17m26s, executable 2m21s, aggregate passed). Native CPU AI, isolated rational and scoped metadata passed. Research-book build failed; deployment skipped. Overall CI is therefore not wholly green. The locally reproduced historical publication-source refusal remains a separate blocker; all independently exercised supported production behavior passed. No merge or deployment performed.