Repository navigation
knowledge: qualify HF Qwen3.5 conv/GDN bindings - #250
Conversation
Trace fixed Transformers releases through module/instance binding, separate conv and recurrence selection, cache phases and classification-prefill limits. Retain source identities, five representative cases and candidate/not_run state. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Preserve the upstream repository-evolution references and the HF source slice before regenerating the native discovery and coverage views. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.com>
Refresh the native generated projections from the committed source union. Keep candidate state and source-derived validation boundaries unchanged. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: Dong1017 <55723162+Dong1017@users.noreply.github.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. |
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 f0dfee5
Independent review of 173e9c3cb56a83fe31cf0c2d3184c429db303781...f0dfee58054c2be8d17afd75d78eb5782a861ff8: no actionable findings in this source-knowledge slice. I checked the complete 85-file diff against the pinned Transformers 5.10.2/5.17.0 sources and bounded worker seams, including constructor versus module-global binding, Hub/worker forward overrides, independent conv/GDN fallback, cache-phase guards, state layout/dtype and conditional classification-prefill reuse. Existing Dao/vLLM material and the three upstream RTK canonical files are preserved.
Verification: all 13 retained upstream files match their SHA-256 and Git blob identities; all 85 published file blobs match the reviewed checkout. The seven HF canonical files are unchanged through integration, and the 77 native index outputs plus coverage match their generation receipts. I inspected the actual Ready CI logs: both Python versions passed lint and 2,465 tests (11 skipped; 259 subtests passed). Both installed-package jobs built a noneditable wheel, installed it in a fresh environment and passed the isolated installed-resource check outside the checkout with PYTHONPATH unset.
Explicit disposition of the skill-evaluation advisory: its 85 uncovered paths are accepted here only as reviewed source/software changes. Zero comparisons were selected and execution was skipped; this is no efficacy, numerical/model, hardware or performance acceptance. The original Issue #141 installation/ready and matched empirical-delta criteria remain open.
The automated Codex review failed to complete; I do not count it as successful review evidence. Fresh REST reviews/inline comments and GraphQL review threads contained no actionable finding to resolve. This COMMENT records my own independent review, not merge or Issue-closure authority.
Independent reviewer: Codex / gpt-6.1-sol [effort=ultra] (review-only; no material authorship).
Governing Issue and delivery slice
Refs #141
kernel-writing-guideowner with the concrete HF Qwen3.5 caller contract.Knowledge authorship declaration
Changed artifacts and evidence
references/hf-qwen35-gdn-binding.md: 5.10.2 saves instance callables; 5.17.0 calls decorated module globals and permits Hub/worker forward replacementreferences/hf-qwen35-gdn-source.jsonb24d3525755376d2b0a5184ce66526b120729e6erender_indexand coveragemaineach completed once; all 77 index outputs and coverage read back against their returned/written bytesThe primary source is one repository at two immutable revisions: Transformers 5.10.2 at
0dad7b822255a0ae261ec45ae937371e859ffd1aand 5.17.0 at856157a2f3e9594954310df18fdccc31ffddebe9. Two fixed existing worker files are bounded caller context only. Original upstream 173e repository-evolution material is retained in the source union.Checks and dependencies
tests/test_knowledge_contract.pyandtests/test_skill_catalog.py: 27 passed, 2 skipped, exit 0 on the source checkpoint. All seven HF canonical files remained byte-identical through upstream integration. Source integrity, local links, real checkout/catalog resource reads and committed-input/native-output guards passed. Working tree is clean.candidate/not_run.Remaining work and Issue closure