Repository navigation
Conversation
|
Important Review skippedToo many files! This PR contains 129 files, which is 29 over the limit of 100. To get a review, reduce the PR to 100 files or fewer by splitting it into smaller PRs or changing its base branch. Upgrade to a paid plan to raise the limit. This review couldn't start because sufficient usage credits or metered capacity aren't available. Add credits or update usage-based reviews in the billing tab, then retry.
 |
3cac638 to
473da4e
Compare
|
@doraemonmj Requested your review: this adds source-bound execution receipts and coverage reports to the onboard environment jobs; 13 pass / 2 skip. Note the PR is currently CONFLICTING, so the author needs to rebase before merge. |
473da4e to
495e41d
Compare
|
Thanks for this — a thorough, well-tested change. I reviewed the full diff (merge-base Should fix
Consider
Not yet demonstrated on hardware
Agree with keeping it draft until default-onboard + SDMA acceptance complete. |
Classify scene tests by product, runtime and execution mode, and verify mandatory selected cases against source-bound parent and child receipts. Preserve device dispatch, exact poison retries and scoped skip exceptions, including existing quarantines and newly collected native consumers. Enable verification in existing default and torch_npu onboard workflows. Name the framework suite by environment across PR, Daily and comment callers, including setup actions, artifacts and documented check migration. Preserve triggers, device leases and SDMA isolation. Retain delegating setup and identity entry points for comment workflows loaded from the default branch while the migration is pending. Pin comment-triggered execution to a verified head/base merge snapshot, so older PR heads consume matching CI interfaces. Reject unavailable or stale merge snapshots before admitting any suite and report all three SHAs. Publish per-feature selected, body-entered, passed, failed, skipped and missing counts with source/environment identity and links to evidence. Keep partial failure evidence without qualifying missing or invalid receipts; failures cannot be hidden by retries or cached green summaries. Keep toolchain bootstrap commands usable before SDK installation and cover their standalone entry points without site-packages. Preserve policy-error inventories as JSON with a nonzero exit, and scope the Daily counting-window PMU skip to its exact platform and reason. Related to hw-native-sys#2904
495e41d to
7b5f08e
Compare
|
@doraemonmj Thanks for the review. Addressed the unused-output finding: removed the Added an assertion that emitted gate outputs match the job's declared outputs. All 46 related comment-workflow tests, Ruff, and actionlint passed. These are focused validation results, not a claim that the updated head's full CI has completed. |
|
@doraemonmj Re-review requested on head |
|
/run-a5 |
1 similar comment
|
/run-a5 |
|
/run-a5 finished without all suites passing Tested head:
|
|
/run-a5 |
|
/run-a5 finished without all suites passing Tested head:
|
|
Two suggestions to keep the coverage check maintainable:
|
|
@zxx0288 The PR is now CONFLICTING after the recent |
Review follow-up (473da4e)
policy_errorJSON and exit nonzero rather than INTERNALERROR. Document and test that auxiliary probes do not qualify Kernel capture gaps.Earlier CI evidence below is historical and does not qualify this updated head. New CI and administrator-controlled A5/Daily validation remain required; no comment-triggered hardware run was posted by this repair. The updated comment definitions only become active on the default branch after merge; candidate-ref Daily workflow_dispatch is needed to validate candidate workflows before then. PR remains draft.
An onboard job's exit status alone does not show which promised features actually ran. This change names the framework suite by its dependency environment and verifies declared feature groups against source-bound execution receipts inside the existing default and torch_npu jobs.
Related to #2904.
Changes
st-framework-<arch>, its reusable workflow to_st-framework.yml, and setup action tosetup-framework. Update PR, Daily, emergency and A5 comment callers, artifact paths, contracts and documentation together. Preserve trigger gates, device leases, manual selection and A2/A3 main/SDMA/fault isolation.Most scene-test/example changes only add metadata; they do not change test bodies or parameter values. No separate capture job or new automatic A5 PR scheduling is introduced.
Validation
ea173476a623533246f7f3c16ec967a4b1fa7b9c; the PR remains one commit.fe215517c955ccc8fff367c99ac9b77695eb80a5report 15 native HBG capture passes and 24 framework eager passes per runtime. This is evidence for the previous revision, not qualification of the updated head.Current-head CI evidence
Run 42559820386 tests PR head
04a921c22b33e99440a4ac05cb269d7cc2561b44at merge source8fb23b3153e7fd5046cdcac54f172a44ac53d364.507011on the first injected failure and507046on the second; device logs includeHandleTaskTimeout. Scenario 13 matches the existing #2923 symptom. This PR changes no runtime/platform implementation; the root-cause repair remains separate. SDMA and fault partitions were not reached after the main step failed and are correctly marked incomplete.403: The workflow run containing this job is already running. No rerun has started. Preserve the first-attempt evidence; a later passing retry would be a non-reproduction, not proof that [Bug] HBG capture second boot-failure replay times out in A3 CI #2923 is fixed.Remaining acceptance
Keep the PR in draft: default onboard acceptance remains blocked by the observed second-fault replay failures, and SDMA partitions still need execution. The renamed framework job and its uploaded evidence have been verified at the current head. Local regression tests use CPU stand-ins and do not qualify hardware execution.
Administrators must confirm any external required-check migration and trigger A5 through the approved
/run-a5policy. Public main rulesets currently show no required status-check rule; the authenticated branch-protection endpoint returned 404 to this write-level account, so legacy protection has not been independently verified or modified. The migration procedure is documented indocs/ci.md.Daily/manual scope and A5 require their own real execution evidence. Torch dtype-dependent MXFP cases are generated before collection and inventories must be compared under the actual runner dependencies. Unsupported A5/TMR/framework capture combinations remain explicit gaps. These summaries do not claim simulation, Network1, standalone model or DFX-smoke coverage, and this PR does not by itself establish complete issue acceptance.