Repository navigation
Conversation
Skill evaluationPLANNED: not executed; there are no task outcomes yet.
Author verification steps (nonpaid; not yet run)
python tools/build_skill_index.py --source-revision 75afe47c07e337e9592010cb67b079777515fd08 --checkThe declared source pin matches this head's canonical library bytes. The command has not run: retain its exit code and full drift/success output. It checks the generated projection, not claim truth or skill efficacy. Evidence and logs · Retention: 30 days. Plan bindings, budget, frozen exam and advisory detailsAdvisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
Model/settings:
Automatic execution / cumulative PR budget: {"reason":"Trusted configuration or pricing missing, invalid or expired","remaining_usd":"unknown","reserved_or_unresolved_usd":"unknown","spent_or_legacy_usd":"unknown","state":"blocked"}Projection check selection (not an executed result): {"declared_in":"src/inferrsi/data/repo_packs/inferrsi/knowledge/README.md","source_revision":"75afe47c07e337e9592010cb67b079777515fd08","status":"ready"}Provider/settings: Uncovered / unsupported paths (86)
|
`methods/performance-ceiling-analysis` L1 divides a count by a rated peak and never revisits the model, so an error in the count reaches the floor unchallenged and no later roofline step re-examines it. The skill had no procedure for checking the demand side; `safetensors` appeared nowhere under this owner while the loading mechanics already lived next door. Add a two-gate prerequisite: the bidirectional per-tensor checkpoint join, and the separate per-rank weight ledger. The first needs no parallelism layout, so passing it certifies nothing about per-device bytes, and the distinction is stated rather than implied. State the blind spot too: the anchor covers the weight traffic term only, since a windowed and a global attention with equal head dimensions have byte-identical weights. Record the reviewed external source it derives from at a pinned commit, with per-file git blob and sha256 digests and per-file inspection scope. Documentation claims and source-verified claims are marked separately. Carry the source's two-threshold rule, which grants rejection authority only to a rated-peak bound and not to a per-operation roofline, and its archived guard-ablation study with the source's own limitations intact. What the source does not establish is recorded as explicitly as what it does: no end-to-end validation of predicted against measured serving performance within the inspected scope, unmeasured collectives upstream, and companion-blog corpus figures that the repository documents do not state and that must not be cited. `locally_verified` stays false; `empirical_validation` stays not_run. Refs: InferMatrixLab#348 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Collaborator
|
Review of the full PR (new PR, no earlier review of mine): base Not approving yet; the content itself is careful (both gates, the weight-only scope limit and the "does not establish" list read well, and I checked that the 60+ generated records outside Blocking / needs action
Provenance questions (please confirm or downgrade the marker)
Minor
Not verified by me: the HuggingArch file hashes/blobs, and the PR's local |
fake0fan
force-pushed
the
docs/ceiling-checkpoint-anchor
branch
from
November 4, 2026 09:28
b5b7dda to
18c9a16
Compare
Collaborator
|
Incremental review What changed: I compared the per-file patches of New at this head
Still open from last round (no change in this delta)
Static review of the delta plus the hosted CI logs; I did not run anything locally. |
Dong1017
added
the
area:knowledge
ChatGPT Codex Connector
The rebase kept only the coverage-view refresh, so the packaged records still pinned the previous main snapshot and the performance-ceiling-analysis projection was stale. Regenerate with `build_skill_index.py --source-revision 75afe47`. Apart from that skill's record, every record changes only its source pin. Agent-assisted-by: Claude Code / claude-opus-5-5 (author) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Chenguang Zheng <zhengcg0505@gmail.com>
Collaborator
|
Incremental review Fixed in this delta
Still open
Static review of the delta plus hosted CI status. I didn't run anything locally. |
6 tasks
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
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.
Refs #348 (the issue's acceptance owner decides closure; a merged PR does not satisfy the goal by itself).
Rebased onto current
main(663f3b32) after the branch's original base went 77 commits stale.main's two added paper sources under this owner (Mooncake, FlexGen) and its added limitations are preserved; this PR appends beside them.What this changes
methods/performance-ceiling-analysisL1 divides a count by a rated peak and never revisits the model, so an error in the count reaches the floor unchallenged and no later roofline step re-examines it. The skill had no procedure for checking the demand side:safetensorsappeared nowhere under this owner, while the loading mechanics already live inweight-loading-optimizationand a byte-level weight inventory already exists as an artifact pattern underrepositories/h3-c/.This adds the missing prerequisite and records the reviewed external source it is derived from.
methods/performance-ceiling-analysis/SKILL.md.../references/huggingarch-checkpoint-anchor.md.../references/huggingarch-source-identity.json.../metadata.jsondocs/reference-matrix.mddocs/skill-coverage.mdtools/build_skill_coverage.pysrc/inferrsi/data/repo_packs/inferrsi/knowledge/**(80 files)tools/build_skill_index.pyat the content commit; only the performance-ceiling-analysis record changes beyond the source pinThe two gates
The blind spot is stated rather than implied: the anchor covers the weight traffic term only. A sliding-window and a global attention with identical head dimensions have byte-identical weights, so KV growth and boundedness are unanchored by it.
Provenance discipline
[code]where verified against source files at that commit and[doc]where only the project's documentation states them.locally_verifiedstaysfalse;empirical_validationstaysnot_run. No measurement, benchmark or performance claim is made or implied.Out of scope
Deliberately excluded, per #348: the bound model proposed in #303 (this is a prerequisite check, not a bound, and does not depend on #303 being accepted); the
Class A/Class Bvocabulary of #303 — note the cited source usesClass A–Class Efor an unrelated verification-rule taxonomy, and the reference document flags that collision; importing or vendoring any measured kernel corpus; any new skill ID orskill-index.jsonchange.CI state at this head
The skill index is regenerated. The rebase had dropped the index refresh, so the packaged records still pinned the previous
mainsnapshot. That broke fourtest_frozen_offline_pilotcases, twotests/test_skill_index.pytests and the clean-wheelSKILL.mdhash check. Commit1694a0b5regenerates it:All 79 records move their source pin from
7aa3344b…to75afe47c…. Onlycommon/skills/performance-ceiling-analysis.jsonchanges beyond the pin: its two new reference files, updated hashes and the HuggingArch source entry.ascend/skills/torch-npu.jsonchanges only its pin. No generated card was hand-edited.The authoring environment has git 2.34.1, which lacks
--no-lazy-fetch. The generator ran through a local wrapper that drops only that flag. This clone has no promisor remote, so no lazy fetch could occur and the hashed git blob bytes are unaffected. CI reruns the same checks with its own git.Local verification at this head on Python 3.11:
ruff check src tests toolspasses.tests/test_skill_index.pyandtests/test_collect_selection.pypass.tests/check_installed.pysmoke passes.tools/build_skill_coverage.py --checkpasses; the coverage view needed no change.pythonexecutable that is not on PATH. The regeneration removes the six index failures and adds none.The knowledge-model policy still needs an exception approval. See the declaration below. The exception scope now also lists the 80 regenerated records, because
src/inferrsi/data/repo_packs/**/knowledge/**is a protected path.Knowledge author declaration
The declared model identity and effort are not in the trusted allowlist, so this requests an exception under
docs/knowledge-model-policy.md. Attribution is truthful and is not mapped onto an approved identity to obtain a green status. The exception needs a top-levelknowledge-policy-approvalcomment from an authorized independent owner; the PR opener cannot grant it, and new commits require renewed approval.docs/skill-coverage.mdis outsideprotected_pathsand so is not in the exception scope.🤖 Generated with Claude Code