Repository navigation
knowledge: reconcile and integrate all pending knowledge updates - #177
Conversation
- Updated the DI calculation guide with revised SLO values for high, medium, and low priority bugs. - Added a new section for medium priority models in the model owner documentation, including a list of models and their respective owners. These changes enhance clarity and provide updated information for better resource allocation and ownership tracking. Signed-off-by: wangyu <410167048@qq.com>
- Updated the 'owners' index with the latest modification date. - Added a new 'module_owner' document to track key modules and their respective owners for better routing during reviews and debugging. - Removed the medium priority models section from 'model_owner' documentation to streamline the content. These changes enhance the clarity and organization of ownership information within the vLLM-Omni project. Signed-off-by: wangyu <410167048@qq.com>
- Removed the high-priority section from 'model_owner' documentation to streamline content. - Updated 'module_owner' documentation with new owner assignments for several modules, including 'Benchmark/metrics/profiler', 'scheduler', 'Engine', and 'api_server'. These changes improve the clarity and accuracy of ownership information within the project. Signed-off-by: wangyu <410167048@qq.com>
- Changed the 'updated' date in the 'model_owner' documentation to reflect the accurate date of August 11, 2026. This update ensures the documentation remains current and reliable for users. Signed-off-by: wangyu <410167048@qq.com>
- Introduced a new document outlining release rules for vLLM-Omni, detailing pre-release requirements and acceptance criteria. - Added a new section for 'release' under the schema to categorize the release-related topics. - Updated the 'updated' date in the vLLM-Omni index to reflect the latest changes. These updates enhance the clarity and organization of release processes within the vLLM-Omni project. Signed-off-by: wangyu <410167048@qq.com>
- Revised the release guidelines to include CUDA L1 in the pass rate requirements, ensuring comprehensive coverage of all relevant CUDA levels. - Updated the NPU pass rate criteria to include L1, enhancing the clarity of performance expectations. These changes improve the accuracy and completeness of the release requirements for the vLLM-Omni project. Signed-off-by: wangyu <410167048@qq.com>
- Adjusted the formatting of the CUDA pass rate requirement for clarity, ensuring consistent spacing in the documentation. - This minor update enhances readability without altering the content or meaning of the release criteria. Signed-off-by: wangyu <410167048@qq.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. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: c302bc8393
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| | 理解调度器继承链、KV transfer 与 prefix cache 语义 | [architecture](architecture.md) | | ||
| | 按 PR 描述直达 prefix cache、token budget、upstream 接口或 side-stream 首批源码 | [rules / Direct 代码快速入口](rules.md#direct-代码快速入口) | | ||
|
|
||
| - [rules-shared-lifecycle](rules-shared-lifecycle.md) 新增核对：SCHED-6a2。 |
There was a problem hiding this comment.
For a review that touches OmniSchedulerMixin, Direct resolves the exact owner route to components/scheduler/rules.md, whose quick map neither mentions SCHED-6a2 nor links this new page; the only route added here is in _index.md, which reviewers are explicitly told not to reopen after exact routing. Consequently the new terminal-output/stats/cleanup checks are not applied. Add this topic to the owner rule's Direct map, or union its unique details into the existing SCHED-3a rule for PR #5461.
AGENTS.md reference: knowledge/AGENTS.md:L16-L16
Useful? React with 👍 / 👎.
| ## COMFY-1a — video mode 与 multipart 字段必须共享当前能力矩阵 | ||
|
|
||
| - 触发：ComfyUI video node/client 修改 frame、image/audio/video references 或 mode。 | ||
| - 强制：先按当前服务端和目标 checkpoint 的能力矩阵校验组合，再确定唯一 mode；node 与 client 使用同一矩阵，按服务端字段名和输入顺序编码 multipart。新增 first/last frame 或 reference 组合必须同步两端。 | ||
| - 禁止：按字段遍历顺序覆盖 mode；把早期 MiniMax H3 的 reference 限制推广到所有模型或后续功能。 | ||
| - 验收：每个当前支持的组合断言最终 mode、multipart key/count/order；相邻非法组合在发送前失败。^[PR #5756] |
There was a problem hiding this comment.
ComfyUI is already assigned to the tooling owner by release_baseline.yaml (apps/ComfyUI-vLLM-Omni/) and PR #5756 is already represented by tooling/rules.md as OMNI-TOOL-1a/1b. Creating a second serving-component rule for the same behavior gives one source path two owners and two overlapping rule bodies, so later capability updates can diverge and Direct routing will continue selecting tooling rather than this page. Merge any new generalization into the existing tooling rules instead.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
| title: "发布规范" | ||
| created: 2026-08-28 | ||
| updated: 2026-09-22 | ||
| type: guide |
There was a problem hiding this comment.
This PR-derived page declares all non-Guide rows to be mandatory release gates, but type: guide and the release-gates.md filename exclude it from KnowledgeCatalog, which only catalogs rules.md/rules-<topic>.md pages with type: rule. Automated rule intake and owner-rule inventory therefore cannot target these gates, even though they are intended to block releases; move the hard requirements into the nearest CI rule page and leave only explanatory procedure in a guide.
AGENTS.md reference: AGENTS.md:L12-L17
Useful? React with 👍 / 👎.
| - 验收：产物能追溯各 arm 的真实 source/config/environment，报告候选差值和控制漂移； | ||
| 稳态统计可从相同筛选重算，正确性另由对应 parity 测试支持。 ^[PR #7648] | ||
|
|
||
| ## BENCH-1f2 — benchmark 结果必须始终导出样本计数，包括零 |
There was a problem hiding this comment.
BENCH-1f2 violates this page's advertised BENCH-数字字母 format and does not match check_review_report.py's stable-rule expression ^[A-Z][A-Z0-9]*-\d+[a-z]?$. A coverage report therefore cannot identify or validate this rule. Assign the next valid stable ID instead; the same digit-after-letter pattern was introduced repeatedly in this batch and should be corrected consistently.
Useful? React with 👍 / 👎.
| - 验收：昂贵 server startup 前完成 revision/digest preflight；exact case 满足 | ||
| [DIFF-3a](../../components/diffusion/rules.md) 与 [VOMNI-CI-1b](../../ci/rules.md)，视觉和 | ||
| 音频 scorer 都有正向结果及损坏单一 modality 的反例。 ^[PR #5709] |
There was a problem hiding this comment.
The VOMNI-CI-1b link points to ci/rules.md, but that heading exists only in the newly added ci/rules-review-validation.md. Following this acceptance dependency therefore opens a page with no matching rule and prevents a MiniMax-H3 reviewer from reaching the required platform-oracle contract; target ../../ci/rules-review-validation.md#vomni-ci-1b-... instead.
Useful? React with 👍 / 👎.
Pending knowledge updates overlap across ten intake batches and three older proposals. Several reuse IDs for different rules, target the wrong model owner, or restore obsolete release facts. This batch reconciles their useful content against current
mainand preserves the newer44d3ae10release baseline.Validation: all 85 cited upstream PRs confirmed merged; both knowledge validators pass with zero errors (31 structural advisories); release audit
enforceis CLEAN with zero issues; 59 targeted release/adapter/routing/knowledge/skill/capability tests pass, with the 36 knowledge/routing tests repeated after ID refinement. The Direct quick-map regression passes; 49 contract/SDK tests pass; doc links/citations and strict SPEC freshness pass. No runtime code changes.Integrates #39, #42, #70, #150, #156, #158, #159, #161, #162, #163, #165, #171 and #175. The explicit ancestry merge follows the content-reconciliation commit and keeps that reviewed tree; it does not restore the original conflicting snapshots. Use a merge commit to preserve those PR histories.