Repository navigation
knowledge: trace vLLM GDN convolution phases and state - #169
Conversation
Trace the Qwen3-Next in-tree convolution caller, effective phase/backend selection and mutable state contract. Add source-only representative cases without asserting an HF binding or runtime qualification. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Generate the packaged catalog from canonical source commit e862b0f using the stock native generator. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Source/content-scope LGTM at head 1d1c1d50185924c0c410472ad8215ad2f0dc0e85, base 3069b6df6794267306e2cde2d754d0960d3b7470. This is not an unconditional merge-ready conclusion.
Reviewed the new reference/cases, metadata, additive matrix rows and generated changes against #141's scope and prior discussion. Direct inspection of pinned vLLM 537a79f0fc5ba564d2bef43d54915b2f80f9f77d supports the mixed-batch convolution/recurrence split, first/resumed and speculative metadata, requested/effective backend distinction, interleaved-layout fused-MTP exclusion, state-row/initialization contracts and converted-input aliasing. No actionable source-fidelity defect found. Attribution and candidate/not_run/not_evaluated limits are retained. The 59 unrelated generated JSON records change source URL/commit only; the final commit's sole parent is canonical source e862b0ff1b97d0dc9bb10572201121bd2b159800, and that final commit changes only packaged projections.
Software verification: all four Python 3.10/3.11 full-check and installed-package jobs in Ready 40452154616, lint, declarations and latest knowledge-model-policy pass. These remote results do not retroactively turn the disclosed local test/build timeouts into passes. I performed static inspection, not a local test run or native/model/hardware experiment.
Outstanding integration prerequisite: Skill Eval publish still fails HTTP 403 on trusted main; plan succeeds and execute is skipped. This is the existing publication problem, not evidence of faulty GDN guidance. The diagnostic reports accepted permission alternatives issues:write or pull_requests:write while job permissions list Issues:write, so expanding permissions is not justified by this result alone. Resolve the operational blocker or obtain an explicit scoped owner disposition before treating the PR as fully merge-ready.
This source-guidance delivery does not complete #141's concrete HF binding, installation/device qualification or verified L20X performance acceptance. No new requirement or model run is requested for this bounded review.
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Integration-status update only at unchanged inspected head 1d1c1d50185924c0c410472ad8215ad2f0dc0e85 (PR API base snapshot 3069b6df6794267306e2cde2d754d0960d3b7470). Actual main is 38d5883b8e72e352865e214194cb22dc99253dbd; GitHub now reports mergeable=false / dirty. The prior source-scope review remains valid historical evidence, but is not current integration clearance.
Before merge, reconcile this branch with current main, preserve both accepted contributions, regenerate affected projections only where source-binding rules require it, and verify packaging plus applicable checks on the resulting head. No conflict-file attribution is made without a merge reproduction; no new content/feature requirement is introduced. This is not a new HTTP403 finding: shared publisher recovery was verified separately. No code execution, branch change, workflow dispatch or merge was performed by this review.
Dong1017
added
the
area:knowledge
ChatGPT Codex Connector
Preserve the accepted causal-conv1d guidance and the unpublished GDN case-resource and return-navigation corrections. Merge typed provenance and retain candidate/not_run boundaries. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Generate the catalog and coverage view from canonical source 80e2382 using unchanged native functions. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Skill evaluationAutomatic execution / cumulative PR budget: Advisory: obtain missing evidenceRetained evidence availability: planned / unexecuted. No approval or merge authorization is issued.
PLANNED: execution is not authorized by knowing this hash. Selected comparisons: 0 · Estimated task calls: 0. Applicable only while these revisions and inputs match the current PR. Plan bindings, budget, frozen exam and advisory detailsUnchanged-correct is valid neutral evidence, not incremental benefit. Per-case outcomes (discovery excluded; regressions and gaps first): []Showing at most 20 of 0 pairs; the evidence link contains all outcomes, including omitted regressions/gaps. Plan SHA-256: Model/settings:
Provider/settings: Uncovered / unsupported paths (85)
|
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM 161bc88
Independent review by a separate Codex agent (gpt-6.1-sol, effort ultra), with no material authorship. Submitted as COMMENT under the shared Dong1017 account.
Reviewed against main 013a481, canonical source 80e2382, and its sole generated child above. No material source-correctness or integration blocker found for this bounded knowledge delivery.
- The pinned vLLM readset (12 files, 443907 bytes) supports the mixed-phase convolution/recurrence split, first/resumed and speculative metadata, requested/effective backend distinction, interleaved-layout fused-MTP exclusion, state initialization/row mapping and converted-input aliasing. Apache/vLLM/Tri Dao/FLA notices remain distinct from the preserved BSD causal-conv1d guidance.
- Earlier main materials and metadata are preserved. The four repaired case/reference resources retain their exact prior blobs. One case record contains four scenarios; inventory is 74 registered skills, 118 SKILL.md documents, 59 local profiles, 5 structured cases and 122 catalog records.
- All 496 committed source hashes and all 76 generated outputs match the captured native returns. Native modules are unchanged; G changes only generated projections. The other 73 generated skill records change only source URL/commit bindings.
- I inspected all four decoded Ready job logs. Python 3.10/3.11 each report 2355 passed, 11 skipped and 183 subtests passed, with Ruff passing. Both isolated installed-wheel checks report 122 records and validate the complete indexed resource bytes. Their checked merge tree equals this G tree. The checkout suite exercises case bundle/link resolution; a separately invoked installed-root case-ID lookup was not run. Lint, declarations and the current knowledge-model-policy status also succeed.
- The current Skill Eval plan/publish run succeeds and publishes plan 7407e553bea35ca8b8be1990bb51d2cc891061663046547e96885ede423f80d6, bound to this head/base/trusted main. Both ZIP digests and the identical plan bytes were verified. It selects 0 comparisons, estimates 0 calls and leaves all 85 changed paths uncovered; execute is skipped, and audited-budget availability remains blocked/unknown.
Explicit reviewer disposition: those advisory gaps are acceptable for source/software delivery only and remain unassessed for efficacy. Retain candidate/not_run/not_evaluated; this review provides no efficacy PASS, model/provider/device execution, numerical or performance qualification, HF binding, sm_89/L20X readiness or whole-caller acceptance. Issue #141 remains open. I did not edit source, rerun generators, install dependencies or execute local tests.
Document the pinned vLLM Qwen3-Next GDN convolution contract: effective convolution and recurrence selection, mixed/speculative metadata, state-row initialization, accepted-token offsets and in-place mutation. The existing kernel-writing-guide remains the owner.
The source reference, four representative source scenarios and typed provenance remain source-derived candidate/not_run guidance; agent efficacy is not_evaluated. The four scenarios share one structured case record. Its declared resources include the case record and the case-local contract, with return navigation preserved from the unpublished repair.
Source: vllm-project/vllm @ 537a79f0fc5ba564d2bef43d54915b2f80f9f77d, tree d573bc73908ae9cf2e1254710483ae3467d12343, originally inspected 2026-10-27. All 12 retained source files (443907 bytes) match their pinned Git blobs. Root Apache-2.0, the convolution file's vLLM/Tri Dao attribution and the separate vendored FLA MIT notice remain distinct. This contribution paraphrases the contract and does not vendor upstream implementation or establish standalone dependency API equivalence.
Integration preserves the complete current main at 013a481, including the accepted BSD-3-Clause causal-conv1d guide/resources, every existing matrix row, older provenance limits and executable file modes. GDN tags, sources, readsets and limits are merged into the existing owner metadata. The original GDN reference is unchanged; the case-local copy differs only in its return-navigation href.
Canonical source commit: 80e2382 (ordered parents: 1d1c1d5, 013a481). Generated-only sole child: 161bc88.
Validation:
No upstream kernels, model/provider calls, device runs, installation or performance experiments were executed. Concrete HF caller binding, physical state allocation/recycling, sm_89/L20X qualification, numerical results, performance and whole-caller acceptance remain separate from this change.
Refs #141