Skip to content

feat: add vllm-project/router issue priority rules - #445

Merged
zengchuang-hw merged 2 commits into
mainfrom
cursor/vllm-router-issue-priority-0606
Nov 10, 2026
Merged

zengchuang-hw merged 2 commits into
mainfrom
cursor/vllm-router-issue-priority-0606

Conversation

@yenuo26

@yenuo26 yenuo26 commented Nov 10, 2026 •

Copy link
Copy Markdown
Collaborator

摘要

把已约定的 vllm-project/router issue 分级规则写入知识树,供 InferMatrixCopilot 在该仓库分诊 issue 时使用。四档为致命、高、中、低。先看影响,再看特性:致命和高按影响判定;中和低按特性给默认档。主路径不可用或大面积 CI 失败时必须升档,特性默认档不能压低。

规则页是原文短文,标题就是「总原则 / 致命 / 高 / 中(默认特性)/ 低(默认特性)/ 判定顺序」,没有 ROUTER-PRI-* 编号,也没有「触发 / 必须 / 禁止 / 验收」模板。校验器不要求这类 ID。

放置

  • 规则只适用于 vllm-project/router,所以新建 canonical slug vllm-router,不放进 general/,也不并入其他仓库。
  • 无关任务不必默认读到这份分级,且已有更窄的工作主题 bug,因此正文在 knowledge/repos/vllm-router/bug/rules.md,不进仓库根 rules.md。
  • 仓库入口只做场景路由。特性例子和四条 GitHub issue 链接保留在中档规则里。
  • playbooks/issue-triage.yaml 仍是仓库中立的,没有写入单一仓库路径。adapters/vllm_router 把分级规则放进该仓库的 briefing,分诊时会读到原文。

校验

  • python knowledge/tools/check_knowledge_tree.py
  • python knowledge/tools/check_wiki_lint.py
  • test/test_adapters.py::test_shipped_vllm_router_adapter_injects_priority_rules
Open in Web Open in Cursor 

cursoragent and others added 2 commits November 10, 2026 09:04
Store the agreed fatal/high/medium/low grading guide in the vllm-router
knowledge slice so issue triage can apply it, including the impact overrides.

Co-authored-by: wangyu <yenuo26@users.noreply.github.com>
Drop ROUTER-PRI IDs and the trigger/must/forbid/accept template so the
page matches the agreed short grading guide.

Co-authored-by: wangyu <yenuo26@users.noreply.github.com>

@KaisennHu KaisennHu left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@KaisennHu

Copy link
Copy Markdown
Contributor

CI 状态:commit 5ebd9bf 上 push 触发的 tests / suite 失败(run 42789925284),失败用例是 test/test_run_shutdown.py::test_module_cancelled_mid_attempt_marks_interrupted_and_reraises,AssertionError: assert 0 == 1(缺 module_interrupted 事件)。

同一 commit 由 pull_request 触发的 run 42789934179 中该 job 通过,且失败点不在本 PR 改动的 adapters/、knowledge/、test/test_adapters.py 范围内,优先判断是 shutdown 相关的 flake。建议先 rerun 失败的 suite,绿了再合;如果 rerun 后仍稳定复现,再单独看 push 与 PR 两次运行的环境差异。

@zengchuang-hw
zengchuang-hw merged commit 49076e8 into main Nov 10, 2026
3 of 4 checks passed
zengchuang-hw pushed a commit that referenced this pull request Nov 10, 2026
## Summary

Extends the `knowledge/repos/vllm-router/` slice, which #445 created and
has since merged, with CI and release knowledge. It also merges the
repository entry so one page routes both the `bug/` and `ci/` work
topics, and brings the shared DI rule in line with the review feedback
on this PR.

Router statements are checked against `main@ddf78218` (2026-11-09) and
against the published channels: PyPI, Docker Hub, and GitHub Releases.

## Contents

1. `_index.md` — merged repository entry. Routes issue triage to `bug/`
and release work to `ci/`, keeps the package, module, and alias mapping,
records the audited `main` revision, and states that no component owners
or `_routes.yaml` exist yet.
2. `ci/_index.md` — CI entry. Describes the four Buildkite lanes with
their real upload gates (non-default branch plus `ready`; default
branch, merge queue, or `merge-test`; default branch with `NIGHTLY=1` or
`nightly-test`), the jobs each lane runs (P/D smoke, kv_aware e2e, gRPC
P/D, GSM8K, gRPC prefix-KV perf, L1 encode perf, ROCm xGMI and RDMA
matrices), the pytest markers, and the hardware presets.
3. `ci/release-gates.md` — release standard:
1. Version sources: `pyproject.toml` is the only value `version-check`
reads, while four files must be bumped together.
2. Published-artifact cross-check: git tags, PyPI versions, Docker Hub
tags, and GitHub Releases side by side, including tags without artifacts
(`v0.1.11`, `v0.1.13`, `v0.1.14`) and the Docker-only `v0.1.16.1`.
3. Hard gates: tag-to-`pyproject.toml` equality; x86_64 and aarch64
manylinux_2_28 builds with `auditwheel`; binary and wheel smoke tests;
the `twine` PyPI chain; the Docker confirmation sequence; nightly image
publication and 14-tag retention.
4. Quality gates and evidence rules: L2 and L3 results; Python coverage
read from the router coverage dashboard, with Rust coverage gated on a
self-built collector agreed at kickoff; per-scenario performance
comparison; requirement completion; remaining DI; critical-issue
ownership.
5. Failure modes and release steps, including manual GitHub Release
authoring because no pipeline step creates releases.
4. `knowledge/general/bug/guides/di-calculation.md` and
`knowledge/general/bug/_index.md` — shared DI rule refresh: DI is no
longer accumulated by days open; breaching the SLO escalates the issue
one severity level and DI takes the new level's base value. Severity
criteria for critical, high, medium, and low are recorded there so the
router page can link instead of restating them.

## Metadata

No additional metadata changes in this diff. The `vllm-router` ownership
tag and the `repos/_index.md` row landed with #445; this PR reuses them
and only merges the entry page.

## Validation

1. `python knowledge/tools/check_knowledge_tree.py` — no errors and no
warnings for the merged entry, the CI pages, or the two updated DI
pages. Three pre-existing errors remain for three `rules.md` pages
elsewhere in the tree on a local Windows autocrlf checkout; their LF
blobs are under the 32768-byte threshold, so these are line-ending
artifacts of the local checkout and are unrelated to this change.
2. `python knowledge/tools/check_wiki_lint.py` — no errors and no
warnings across 1531 pages.

## Boundary

The vLLM Router slice covers issue triage, CI, and release knowledge
only. `components/`, `models/`, and `_routes.yaml` are not established,
and `adapters/vllm_router` only injects the triage briefing, so the
release page must not be treated as a code owner during review.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

4 participants