Fail closed on Puzzletron correctness failures - #2114
Open
j-rausch wants to merge 6 commits into
Open
Conversation
Treat sort and slicing-equivalence mismatches as blocking correctness errors while keeping ranking-quality findings advisory. Propagate the same contract through parent sweeps and progress reports. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
Explain how width-ranking quality checks differ from sort and slicing equivalence failures, including what is measured and how strict qualification can promote warnings. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
Keep the main README focused on the first-run workflow while retaining the detailed ranking, slicing, and qualification reference in a linked topic guide. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Enterprise Run ID: 📒 Files selected for processing (2)
🚧 Files skipped from review as they are similar to previous changes (1)
📝 WalkthroughWalkthroughSanity handling separates blocking sort and slicing correctness failures from advisory width-ranking warnings. Verdicts, diagnostics, campaign reports, tests, and Puzzletron documentation reflect the new severity and qualification behavior. ChangesSanity policy and execution
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant ParentSweep
participant SanityVerdict
participant CampaignReport
ParentSweep->>SanityVerdict: combine width, sort, and slicing findings
SanityVerdict->>SanityVerdict: classify warnings and blocking errors
SanityVerdict->>CampaignReport: publish stage verdict and finding severities
CampaignReport->>CampaignReport: render gate badges and DAG status
Suggested reviewers: 🚥 Pre-merge checks | ✅ 5 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (5 passed)
✨ Finishing Touches 💡 1📝 Generate docstrings 💡
🧪 Generate unit tests (beta)
Comment |
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## feature/puzzletron_v2 #2114 +/- ##
=========================================================
+ Coverage 54.76% 55.22% +0.46%
=========================================================
Files 704 704
Lines 90869 90907 +38
=========================================================
+ Hits 49761 50200 +439
+ Misses 41108 40707 -401
Flags with carried forward coverage won't be shown. Click here to find out more. ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
j-rausch
marked this pull request as ready for review
August 7, 2026 18:44
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Warning
CodeRabbit couldn't request changes on this pull request because it doesn't have sufficient GitHub permissions.
Please grant CodeRabbit Pull requests: Read and write permission and re-run the review.
Actionable comments posted: 4
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@examples/puzzletron/docs/sanity_validation.md`:
- Around line 39-44: Standardize the deterministic control terminology across
all listed documentation sites: in examples/puzzletron/docs/sanity_validation.md
lines 39-44, choose one canonical label and identify random as a legacy alias
only where needed; in examples/puzzletron/README.md lines 484-503, clarify that
random denotes the original or unsorted prefix; in
examples/puzzletron/docs/v2_architecture.md line 142 and lines 361-386, replace
inconsistent control names and align gate and qualification terminology with the
canonical label.
In `@modelopt/torch/puzzletron/diagnostics/campaign_progress_report.py`:
- Around line 503-504: Update the later empty-state return for non-empty rows
with empty axes or metrics to prefix its message with gate_summary, matching the
not-rows branch. Preserve the existing empty-state content while ensuring failed
slicing-equivalence gates remain visible.
In `@modelopt/torch/puzzletron/stages/diagnostics.py`:
- Line 3343: Unify the post-barrier completion flow around the shared verdict in
the diagnostics stage: after rank 0 evaluates the sort result, broadcast or
otherwise derive the same passed/failed value on every rank, and ensure all
ranks either follow the same failure return path or use rank 0 exclusively for
persisting the stage manifest. Update the `verdict` construction and surrounding
rank-specific handling so a failed result cannot be overwritten by another
rank’s `SanityVerdict(passed=True)`.
- Line 1615: Move the SanityVerdict import at
modelopt/torch/puzzletron/stages/diagnostics.py:1615 and the
complete_sanity_stage import at
modelopt/torch/puzzletron/stages/diagnostics.py:2741-2743 into the module-level
import section. Remove both local imports while preserving their existing
usages.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 642ff8c6-994d-42ec-8334-e1b25f5a47a2
📒 Files selected for processing (15)
examples/puzzletron/README.mdexamples/puzzletron/docs/sanity_validation.mdexamples/puzzletron/docs/v2_architecture.mdmodelopt/torch/puzzletron/diagnostics/campaign_findings.pymodelopt/torch/puzzletron/diagnostics/campaign_progress_report.pymodelopt/torch/puzzletron/diagnostics/sanity_verdict.pymodelopt/torch/puzzletron/diagnostics/width_sanity.pymodelopt/torch/puzzletron/stages/diagnostics.pytests/unit/torch/puzzletron/test_campaign_findings.pytests/unit/torch/puzzletron/test_campaign_progress_report.pytests/unit/torch/puzzletron/test_diagnostic_scoring_config.pytests/unit/torch/puzzletron/test_hidden_width_diagnostic.pytests/unit/torch/puzzletron/test_sanity_verdict.pytests/unit/torch/puzzletron/test_width_sanity_aggregation.pytests/unit/torch/puzzletron/test_width_slice_equivalence.py
💤 Files with no reviewable changes (1)
- tests/unit/torch/puzzletron/test_diagnostic_scoring_config.py
Make every rank consume the persisted sort verdict, keep empty-state gates visible, and distinguish the deterministic original-order control from seeded random permutations. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
Finalize summary evidence and stage manifests before process-group teardown so one rank writes the result and every rank consumes the same completed verdict. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
Contributor
There was a problem hiding this comment.
Warning
CodeRabbit couldn't request changes on this pull request because it doesn't have sufficient GitHub permissions.
Please grant CodeRabbit Pull requests: Read and write permission and re-run the review.
Actionable comments posted: 1
🧹 Nitpick comments (1)
tests/unit/torch/puzzletron/test_diagnostic_scoring_config.py (1)
125-128: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winAssert the blocking contract this test names.
The test is named
..._records_blocking_drift, but it does not check the two fields that carry the blocking semantics: the finding severity and the summary verdict._write_sort_equivalence_summarysetsseverity="error"and"verdict": "failed", and_complete_sort_equivalence_stageconsumes those values. A regression that downgrades the severity to"warning"would still pass this test.💚 Proposed additional assertions
summary = json.loads(summary_path.read_text(encoding="utf-8")) assert summary["passed"] is False assert summary["delta"] == 0.25 assert summary["findings"][0]["stage"] == "sort_sanity" + assert summary["findings"][0]["severity"] == "error" + assert summary["verdict"] == "failed"🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@tests/unit/torch/puzzletron/test_diagnostic_scoring_config.py` around lines 125 - 128, Extend the assertions in the diagnostic scoring test around _write_sort_equivalence_summary to verify the blocking contract: assert the finding severity is "error" and the summary verdict is "failed". Keep the existing passed, delta, and stage assertions unchanged.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@modelopt/torch/puzzletron/stages/diagnostics.py`:
- Around line 3349-3365: Update _finalize_sort_equivalence_stage so the master
rank catches any exception from _write_sort_equivalence_summary, records the
failure state, and still reaches dist.barrier(). After the barrier, propagate
the captured error on every rank so all processes fail consistently without
stranding non-master ranks.
---
Nitpick comments:
In `@tests/unit/torch/puzzletron/test_diagnostic_scoring_config.py`:
- Around line 125-128: Extend the assertions in the diagnostic scoring test
around _write_sort_equivalence_summary to verify the blocking contract: assert
the finding severity is "error" and the summary verdict is "failed". Keep the
existing passed, delta, and stage assertions unchanged.
🪄 Autofix
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Enterprise
Run ID: 8e7e3206-c631-4b18-9de7-e1aabca23dbd
📒 Files selected for processing (7)
examples/puzzletron/README.mdexamples/puzzletron/docs/sanity_validation.mdexamples/puzzletron/docs/v2_architecture.mdmodelopt/torch/puzzletron/diagnostics/campaign_progress_report.pymodelopt/torch/puzzletron/stages/diagnostics.pytests/unit/torch/puzzletron/test_campaign_progress_report.pytests/unit/torch/puzzletron/test_diagnostic_scoring_config.py
🚧 Files skipped from review as they are similar to previous changes (5)
- examples/puzzletron/docs/sanity_validation.md
- tests/unit/torch/puzzletron/test_campaign_progress_report.py
- examples/puzzletron/README.md
- examples/puzzletron/docs/v2_architecture.md
- modelopt/torch/puzzletron/diagnostics/campaign_progress_report.py
Broadcast rank-zero write errors before consuming the summary so all ranks exit consistently instead of hanging at a barrier. Signed-off-by: Johannes Rausch <jrausch@nvidia.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
Type of change: Bug fix and documentation.
Puzzletron sanity stages could record
passed: falsewhile still returning a successful stage. That allowed the campaign DAG to advance after full-width sorting or dynamic-versus-physical slicing had failed a correctness check. Reports also described width-ranking findings as advisory without explaining what they measured or how stage completion differs from campaign qualification.This change:
sanity.fail_on_warningsavailable when a campaign requires warnings to fail the stage;The outcome is that downstream stages no longer proceed using evidence from a correctness-invalid model transformation. Users can also distinguish a weak importance ranking from a broken sorting or slicing implementation and apply a stricter warning policy when scientific or release qualification requires it.
Campaign qualification remains an explicit policy decision. This change does not add a separate
qualification_blockedverdict or prescribe universal acceptance thresholds for ranking quality.Usage
Sort and slicing correctness failures always fail their stages. To also fail a sanity stage on ranking-quality warnings, enable strict warning handling:
Testing
Unit coverage exercises sanity-verdict completion, distributed sort-verdict propagation, warning escalation, parent-sweep propagation, width and slicing aggregation, and report rendering for ranking-quality findings, blocking equivalence failures, and no-metric empty states.