Repository navigation
feat: distill recurrent KV finalization and retirement - #121
Conversation
Trace synchronous LAST_EXITED propagation, later host metadata consumers, and request retirement from fixed vLLM-RLT source. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Generate the shared knowledge projection from committed recurrent KV finalization guidance. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Remove phantom terminal lines from source counts and four end-of-file ranges. Preserve knowledge rules and source bytes. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Bind all shared knowledge projection records to the corrected committed source ledger. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
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.
LGTM for the bounded source-derived candidate scope at 863f680a0747cd86acb9200feb773399178ddb34 (base 48e0d5894117be7c57f8b4ce793c0ec986f219e8). No actionable findings.
Pinned scheduler/finalization, LAST_EXITED copying versus SHARED, host-prefix validation, runner-event chaining and normal-release paths support the documented invariants without conflating copied contents, bookkeeping and device completion. Excluded asynchronous/device-metadata paths remain excluded. Common discovery, six packaged owner resources and candidate/not_run metadata are consistent; retained LICENSE/NOTICE match upstream Git blobs.
Ready CI has all four full-check/clean installed-package jobs successful on Python 3.10/3.11; current knowledge-model-policy status is successful. These are inspected GitHub CI results, not independently rerun tests.
Review covered diffs, owners/metadata, selected immutable source anchors, license identities, inventory and package-resource checks. No native, CUDA/NCCL, model, quality/performance or agent-efficacy experiment was run, nor were all generated hashes independently recalculated. Candidate/not_run is the appropriate retained status.
Integration note: this branch correctly adds one owner to its 58-ID base. When combining parallel knowledge PRs, preserve every owner/matrix update, reconcile aggregate/backend counts, and regenerate shared projections from the actual combined source commit. The other independent PRs being absent here is not a current defect. #91 stays a living tracker; this review is not its acceptance or permission to close it.
Dong1017
added
the
area:knowledge
ChatGPT Codex Connector
Preserve the current catalog and all owner navigation before refreshing the combined source bindings. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Preserve source sets and English code comments before regenerating projections. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Preserve canonical source and validation boundaries. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=xhigh] (author)
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM for the bounded integration delta at 04ca1cde1ae7ef2e6808837dc925d80f2a5d3dc2, reported base 65147a03e465dc838770474736a620e04d72337e. No actionable findings beyond the stated verification limits.
All six RLT owner resources retain reviewed Git blob identities. Current-main canonical content, including SGLang navigation, remains intact outside scoped additions/count updates. Installed-package assertions retain AFD and opt-in selection/context coverage alongside RLT resource, candidate/not_run and LICENSE/NOTICE checks. Counts match one added common owner.
Combined source fc0fda555bb7552f0aac43b2e8831240693a2367 → head changes only generated files: 59 existing records change only provenance and registry hash, while RLT retains reviewed semantics. GitHub reports mergeable.
All four exact-head Ready full-check/clean-wheel jobs pass on Python 3.10/3.11; per-head knowledge-model-policy succeeds. This does not clear the separate aggregate/history failure tracked in #140.
Read-only prior-review/delta/Git-identity/projection/test-diff and CI inspection. No PR code, tests, generator or native/model experiments executed, no exhaustive source-hash recomputation and no new backend/quality/performance/agent-efficacy qualification. Previous candidate/not_run and hardware-coverage limitations remain; issue acceptance and merging are separate.
Preserve main Hopper knowledge, original recurrent KV candidate evidence and both source histories. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author)
Build shared skill records from the actual combined Hopper and recurrent KV source commit. Signed-off-by: GUOGUO <xwdong1998@163.com> Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author)
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Scoped LGTM for the integration delta at 2971186b310c8aff699941ea28115f27c0668c44, reported base 5b939a95d7478b1d8bff013c477972903b7de9ba, compared with previously inspected 04ca1cde1ae7ef2e6808837dc925d80f2a5d3dc2. No new actionable source/integration finding.
All six RLT owner blobs and the original construction report are unchanged. Current-main content is preserved; registry adds only RLT. Installed checks retain MLX/Hopper/AFD and selection/context assertions, with counts reconciled to 109 catalog / 61 registered / 82 SKILL documents. Combined source 876e45a to head is projection-only; other records change provenance only.
Exact-head Ready CI has four successful full-check/installed-package jobs; lint also passes. Fresh per-head policy: success, source https://github.com/InferMatrixLab/InferRSI/actions/runs/40271581200. Aggregate declarations failures/cancellations are separate and are not claimed green. Head and mergeability were refreshed before posting (mergeable=true); recheck required gates and the resulting integration tree before any merge.
Candidate/not_run and separate source-versus-efficacy qualification remain unchanged.
Read-only incremental source/blob/projection/caller/test-assertion and existing discussion/CI inspection. No PR-code execution, generator rerun, model/GPU experiment, thread resolution or Issue acceptance. This records the new inspected head without reopening resolved findings or repeating the upstream audit.
Reviewer: dot / OpenAI assistant; exact runtime model identifier is not exposed.
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM within this PR's accepted source/software scope. Meaningful CI-blocker clearance on unchanged head 2971186b310c8aff699941ea28115f27c0668c44: Ready CI and declarations rerun are successful; exact-head knowledge-model-policy is success. This updates the prior scoped review rather than reopening resolved findings.
Integration inspected against actual main db82ec8346364667b90f4fa97fdb190150141cf0, separately from the PR API's base snapshot 5b939a95d7478b1d8bff013c477972903b7de9ba. The intervening #154/#118 changes (18 paths) do not overlap this contribution. New evaluation-planner mapped canonical directories are not changed by this PR; its source delivery is not thereby covered by a real efficacy evaluation. No new actionable integration finding or unresolved review blocker was identified.
Verification limits: incremental source/integration/prior-thread and CI inspection, no new local tests or model/hardware execution. Existing source-fidelity/packaging conclusions retain their original limits; this is not goal/campaign acceptance or issue closure. Recheck exact head, applicable checks and integration immediately before any merge, especially after another knowledge PR changes shared matrices/projections. Historical red runs remain historical, not rewritten evidence.
Preserve both canonical knowledge histories before rebuilding generated records. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Build the native projection from canonical source 0115f1b. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM for the repaired source/package integration at head 3254e9a25f435c91961e2c8b20943ca5626a6dd0, incrementally compared with previously inspected 2971186b310c8aff699941ea28115f27c0668c44.
All six RLT owner resources and the construction report remain unchanged. Main's Diffusers consolidation and existing Hopper/MLX/Ascend content are preserved. Committed source 0115f1be0539b704b393cffd97598d8bb076bd0c → head is generated-only; unrelated projections change provenance/registry hashes, not content. The previous merge-conflict blocker is cleared (mergeable=true). Independently checked source Diffusers procedure/registry hashes against the projections; affected installed-resource coverage remains intact.
All four Ready jobs and current declaration compliance pass. Skill Eval publisher failed; its operational follow-up remains separate from this unchanged candidate distillation. No new finding and no reopening of resolved concerns.
Integration boundary: inspected actual main 0c78a7e257dfaa3573b77581a39539e8a3a5da4c; the API base remains a4a0fbd62defe2bc53ab090e1fa50bd7139663a6. Their difference is four disjoint Flare #155 paths. This is a scoped source/software conclusion, not an all-workflows-green, model/hardware efficacy, issue acceptance or merge authorization. Separate Skill Eval publication remains failed and execution skipped; do not treat those as a successful evaluation. No model/device calls, installation or workflow dispatch was performed.
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Integration-status update only at unchanged inspected head 3254e9a25f435c91961e2c8b20943ca5626a6dd0 (PR API base snapshot a4a0fbd62defe2bc53ab090e1fa50bd7139663a6). 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.
Preserve the accepted canonical sources and all original candidate resources, merge their registry and matrix contributions, and reconcile aggregate counts. Regenerate projections in a separate child from this committed source. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Build the unchanged native projection from committed source e9d26f1. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Skill evaluationPLANNED: 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. Evidence and logs · Retention: 30 days. Pipeline evidence only; offline fixed outputs do not establish skill efficacy. Plan bindings, budget and frozen examPlan SHA-256: Model/settings:
Uncovered / unsupported paths (84)
|
Omit the stored empty reference list to follow the accepted reader-default contract. Retain all source, applicability, maturity and procedure content. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Build the unchanged native projection from committed source 51660a7. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
Scoped LGTM at 73cf73e5a2c48a5bfbb4725bc6dd8092679a4442, incremental review from 3254e9a25f435c91961e2c8b20943ca5626a6dd0, base 51fa1d31cc4a805490ddfe8055a36c81d70c4f08.
The original substantive owner resources and distillation report are byte-identical. Sidecar change only removes empty references: [], whose reader default remains empty; no source/validation evidence was lost. Regenerated owner projections preserve semantics, unrelated records change provenance only, and current-main content is retained. Prior conflict/integration prerequisite is cleared.
flowchart LR
A[Previously reviewed source contracts] --> B[Current-main integration]
B --> C[Projection and package checks]
C --> D[Bounded software/source clearance]
Ready 40587276119 has all four full-check/installed-package Python 3.10/3.11 jobs successful; exact-head knowledge-model-policy 40606014358 succeeds. A plan with zero selected Skill Eval comparisons is not efficacy evidence.
Read-only source/delta, affected reader/default/projection tests and remote CI inspection. No local execution, model/hardware experiments or new upstream source revalidation. Original candidate/not-run, source-fidelity and host/device limits remain; this is not goal acceptance or automatic merging.
|
Merge-wave integration prerequisite at inspected head #153 changed the same generated catalog provenance records. Integrate current main while preserving both contributions, then regenerate via the existing generator rather than choosing ours/theirs wholesale. Re-run applicable Ready/package checks on the resulting head. This is integration repair, not a new content or efficacy requirement. Existing green checks predate this combined state. flowchart LR
A[Reviewed contribution] --> B[Integrate current main]
B --> C[Regenerate and verify]
C --> D[Merge]
|
Preserve the RLT candidate, current main owners and every matrix addition. Reconcile exact aggregate counts, retaining native reader defaults and separating regenerated discovery from this committed canonical source. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Generate the unchanged native projection from committed source d8a54eb. Preserve current main and original candidate semantics with actual source hashes. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Preserve the complete RLT/current-main union and byte-identical native index inputs and outputs. Refresh the new coverage view separately from this committed integration source. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Use the unchanged current-main coverage generator and its native check. Retain the previously verified index and its real immutable source snapshot. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Expect all 57 normalized profiles after the accepted Ascend owner and RLT candidate are present. Preserve every field, ownership and readset assertion. Agent-assisted-by: Codex / gpt-6.1-sol [effort=max] (author) Signed-off-by: GUOGUO <xwdong1998@163.com>
Dong1017
left a comment
Dong1017
left a comment
There was a problem hiding this comment.
LGTM 1bdae0b
Independent review of the conflict repair and integration against current main 5bea15760982277925e33c0e8deef85035473012. The six RLT owner resources and original construction report are unchanged; current-main owners and maintained matrix rows are retained. Counts reconcile to 66 registered skills, 96 SKILL documents, 114 catalog records and 57 normalized profiles. The one-line profile-count correction preserves every ownership/readset assertion. No actionable findings remain.
Independently verified the complete 67-file index output set and all 351 Git-byte source bindings to d8a54eb8414f29a5152040a7c6bfc5a8f4d09332. Later integration retains those input/output bytes, so this immutable provenance remains valid. The refreshed coverage digest binds all 132 final inputs. Inspected the actual single-render/readback and complete 422-resource package evidence; the RLT candidate/not_run state is preserved.
All four current-head Ready full-suite/clean-wheel jobs pass on Python 3.10/3.11. The inspected 3.11 log checks out this head merged into 5bea15760982277925e33c0e8deef85035473012 and reports 1413 passed, 9 skipped and 150 subtests passed, including the unchanged native index and coverage comparisons. Lint, fresh declarations, current policy and Skill Eval plan/publication attempt 2 succeed. Earlier API-limit failures remain historical. Execute is correctly skipped for the automatic planning event; zero selected comparisons establish no efficacy result.
This clears the bounded source/integration and software-packaging scope. Upstream runtime, device/model, quality, performance and agent efficacy remain unvalidated; Issue acceptance is separate.
Reviewer: Codex / gpt-6.1-sol [effort=ultra], actual independent reviewer.
Refs #91.
Retain the recurrent KV finalization and retirement candidate from
InferMatrixLab/vllm-rlt@4a350344027960a48384626aa646f1765fda6d4f, including synchronous native Ouro / host-KVMetadata scope, source pins/readsets, license/notices, inspection dates and candidate / empiricalnot_runstate. All six current owner resources and the construction report remain byte-identical to the previously published head. The sidecar continues to omit the reader defaultreferences: []; missing references still read as an empty list.Index source
d8a54eb8414f29a5152040a7c6bfc5a8f4d09332first reconciles published73cf73e5a2c48a5bfbb4725bc6dd8092679a4442with main0f85fa8f86b0139167d54785eeda9e038d30859e. Its sole native generated child400288b8bb778dd74d09067fb0882839250bec0acontains 67 index files from one unchanged native render, real revision/input-set/whole-library guards and same-invocation readback. All 351 actual Git hashes and 66 complete record semantics passed. Counts are 66 registered skills, 96 Skill documents and 114 catalogue records; hardware filters are Ascend 62, CUDA 52 and CPU 48. Existing assertions retain precise count updates. Both maintained matrices preserve all main bytes except the original RLT intake-to-candidate row replacement, including accepted LoopX and new Ascend material.Integration source
13b565198002c323e31bf6d5445625b13c9e2c4fmerges that generated child with frozen integration main70adb9f0a963756d040100a25cc708787efa249e, in that parent order, preserving the new coverage-view and policy changes. The complete 422-file library and all 67 index files remain byte-identical. Real native hashes of all 351 inputs at the new integration source equal the index source; its immutable provenance remains valid under the native input-equality contract. No second skill-index render was needed. The subsequent5bea15760982277925e33c0e8deef85035473012main change affects only admission-policy code, tests and documentation; it changes no skill-index or coverage input. Current integration CI must cover the latest base.Coverage head
8819925e33d475aa04ac247e43d76547aa2c3547(treeb0e16e0a96c8400410fa2b035c48a5699c41d391) is the sole coverage-only child of13b565198002c323e31bf6d5445625b13c9e2c4f, changing justdocs/skill-coverage.md. The unchanged current-main coverage CLI and matching--checkpassed. Its input digest issha256:3628f3636b89dd59afd7e8bed3f19f9931adda75336833588bb0def868e7a63b; all 66 registered rows appear once and the other 65 main rows remain unchanged. The native package read at that coverage head matched all 422 committed library resources plus its record, retaining bundle revisionsha256:d61054bf37ef13abdb03d7732c61cadcf960f02596d27898e69c557c3b5a82d4and the RLT candidate/not_run/default-omission state.Current head
1bdae0b4248a87ff5d2dfb3328bb008f443fbde3is the sole test-only child of8819925e33d475aa04ac247e43d76547aa2c3547. It changes onlytests/test_knowledge_contract.py's exact normalized-profile count from 56 to 57, accounting for both the accepted Ascend owner and RLT. Every field, ownership and readset assertion is unchanged. This test file is outside the index and coverage inputs; the complete library, all 67 index files, coverage bytes and their recorded source provenance remain byte-identical to8819925e33d475aa04ac247e43d76547aa2c3547. The previously failing targeted test passed locally at1bdae0b4248a87ff5d2dfb3328bb008f443fbde3. Fresh current-head CI remains required.Four focused local tests passed at
400288b8bb778dd74d09067fb0882839250bec0a; they remain evidence for that head. No full pytest suite or full installed-package check is claimed for this new head. Fresh independent exact-head review and applicable current-head Linux full-render / full-check / clean installed-package Ready CI remain required before merge. Historical Windows CLI/WSL failures and old-head timeouts remain retained. This integration adds no upstream runtime, device/model, quality, performance or agent-efficacy result.