Repository navigation
Conversation
Signed-off-by: hongkuanz <hongkuanz@nvidia.com>
|
Note Currently processing new changes in this PR. This may take a few minutes, please wait... ⚙️ Run configuration
📒 Files selected for processing: 1
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📜 Recent review details
📝 Summary
Merge Risk: ⚪ Minimal · up to
|
| Check name | Status | Explanation | Resolution |
|---|---|---|---|
| Description check |  |
The description clearly states the behavior change, root cause, risk, validation results, compatibility constraints, and related work. It does not follow the required template structure and omits the … | Add the missing template sections: Review map with riskiest files and contract impact; Evidence with explicit Fast CI, Full CI, CodeRabbit reviewed commit, and required Codex reviewed commit; Modeling or data provenance with N/A if not appl… |
-
Check name Status Explanation Title check The title accurately identifies the GLM SM90 KV rejection classification change. It is specific and directly related to the main behavioral change, although it does not name FRAMEWORK_INCOMPATIBLE or … Linked Issues check Check skipped because no linked issues were found for this pull request. Out of Scope Changes check Check skipped because no linked issues were found for this pull request. Cross-Layer Contract PASS. The PR changes only failure classification; it adds no input, schema field, default, or serialization field. The new FRAMEWORK_INCOMPATIBLEpath is called only for build-stage matching errors …Modeling And Data Evidence The PR changes failure-selection logic, not formulas, model data, or performance outputs. The new rule uses an exact typed, build-stage diagnostic match (fpe_support_matrix.py:550-563). Regression tes… Compatibility Boundaries No compatibility boundary changed. The PR modifies only two support-matrix tests and the Python classifier. The classifier adds a build-stage mapping to the existing FRAMEWORK_INCOMPATIBLE status; it … Review Evidence The PR description names the pytest command and result, pre-commit/Ruff/format checks, and the real native replay. It includes boundary validation for wrapped errors, wrong types, altered or incomplet… Full details: Title check
-
Explanation
The title accurately identifies the GLM SM90 KV rejection classification change. It is specific and directly related to the main behavioral change, although it does not name FRAMEWORK_INCOMPATIBLE or FP8 explicitly.
Full details: Cross-Layer Contract
-
Explanation
PASS. The PR changes only failure classification; it adds no input, schema field, default, or serialization field. The new
FRAMEWORK_INCOMPATIBLEpath is called only for build-stage matching errors (fpe_support_matrix.py:550-577, 707-752). Existing JSON/CSV serialization carries the unchangedstatusfield (fpe_support_matrix.py:193-224, 897-905). The qualifier already accepts the status and still rejects it for required probes (qualify_fpe_support_matrix.py:33-42, 98-129). The web rollup and Pages publisher already preserve and render this status (build_fpe_support_matrix.py:139-145, prepare_fpe_pages.py:26-35, 130-144). Tests cover direct and wrapped errors, query-stage failures, altered diagnostics, native H100/H200 and vLLM/SGLang cases, all probe phases, identity, and qualification behavior (test_fpe_support_matrix.py:477-529; test_fpe_qualification.py:88-169). The Rust producer emits the exact diagnostic matched by the classifier (crates/core/src/perfmodel/engine/readiness.rs:101-116).
Full details: Modeling And Data Evidence
-
Explanation
The PR changes failure-selection logic, not formulas, model data, or performance outputs. The new rule uses an exact typed, build-stage diagnostic match (fpe_support_matrix.py:550-563). Regression tests cover wrapped errors, query-stage errors, wrong types, altered messages, altered SM, all four probe phases, H100/H200, and both backend/version pairs (test_fpe_support_matrix.py:477-529). The supplied validation provides provenance and reproducible before/after evidence: nightly run 39922445918, staged source SHA, replay of 96 real plans producing 328 rows, and a field-level comparison showing only status changed. It also reports anomaly checks for unchanged PASS rows and qualification sentinels.
Full details: Compatibility Boundaries
-
Explanation
No compatibility boundary changed. The PR modifies only two support-matrix tests and the Python classifier. The classifier adds a build-stage mapping to the existing FRAMEWORK_INCOMPATIBLE status; it does not change result fields, schema version, function signatures, or defaults. The status and schema version were already present in the base revision and remain accepted by the qualifier. Cargo.toml and python/aisimulate/pyproject.toml are unchanged, including version 0.13.0, maturin binding settings, and the package layout. The changed module adds no Dynamo dependency or binding change, so the one-wheel/one-crate contract and standalone boundary remain intact.
Full details: Review Evidence
-
Explanation
The PR description names the pytest command and result, pre-commit/Ruff/format checks, and the real native replay. It includes boundary validation for wrapped errors, wrong types, altered or incomplete messages, query-stage failures, adversarial diagnostics, and BF16 controls. It distinguishes local and synthetic-fixture checks from real-plan parity results and states that no new nightly qualification or publication occurred. The changed files contain no workflow changes, and the description does not cite inactive nested workflows as hosted CI evidence.
-
Full details: Description check
-
Explanation
The description clearly states the behavior change, root cause, risk, validation results, compatibility constraints, and related work. It does not follow the required template structure and omits the required Codex reviewed commit for a medium-risk change, plus explicit Review map, Modeling or data provenance, and Tracking fields.
Resolution
Add the missing template sections: Review map with riskiest files and contract impact; Evidence with explicit Fast CI, Full CI, CodeRabbit reviewed commit, and required Codex reviewed commit; Modeling or data provenance with N/A if not applicable; and Tracking with Closes and Related PRs. Preserve the existing validation details and state the exact review commit values.
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Comment @coderabbitai help to get the list of available commands.
Signed-off-by: hongkuanz <hongkuanz@nvidia.com>
Summary
FRAMEWORK_INCOMPATIBLE, only for a build-stageValueErrorwith the complete recognized diagnostic.probe_plan, plus wrapped errors, wrong types, changed messages, query failures and required-sentinel rejection.Root cause
Main nightly 39922445918 built/staged the artifacts but stopped during FPE qualification. All 328 unexpected rows were the intentional native rejection of FP8 sparse-MLA KV on H100/H200 with the currently modeled older vLLM/SGLang runtimes. The classifier did not recognize that diagnostic and emitted
BUILD_FAILED.This is a runtime-compatibility category, not hardware FP8 incapability. Unknown build failures and query-time failures remain blocking. A
FRAMEWORK_INCOMPATIBLErow remains nonpassing and cannot satisfy a required known-good probe.Validation
BUILD_FAILEDinstead of the expected compatibility category.PYTHONPATH=python/aisimulate python -m pytest -c python/aisimulate/pytest.ini python/aisimulate/tests/unit/support_matrix tests/test_release_fpe.py -q: 292 passed. The synthetic Git fixture used a process-localcommit.gpgsign=falseoverride; actual commits remain signed.0.13.0.dev202610230000000080, sourceaeb05094a243dc0a054731ac11e63d1af49b6e89; only the checked-out support-matrix tooling was changed.EngineHandle.compileplans, regenerating all 328 originally blocking phase rows. Every field except status matched the original report exactly, including the error and reproducer. No status-only synthetic replacement was used.PERF_DATA_MISSING, not framework-incompatible.buildmarker from the unit regression to comply with CI shard policy; all six new cases remain collected and pass under-m unit.git diff --checkpassed.21618d24686a9963b966ad30e03cc80c785c8a7a; no actionable findings. GitHub CI is running on this head; no remote CI success claim yet.This is local diagnostic validation, not a completed new release qualification or GPU accuracy claim. No nightly rerun/publication was triggered. The fix must land and a fresh nightly must complete before downstream dependency pins advance.
Coordination