Skip to content

docs(knowledge): add vLLM-Router release standard - #444

Open
KaisennHu wants to merge 1 commit into
InferMatrixLab:mainfrom
KaisennHu:knowledge/vllm-router-release-standard
Open

KaisennHu wants to merge 1 commit into
InferMatrixLab:mainfrom
KaisennHu:knowledge/vllm-router-release-standard

Conversation

@KaisennHu

@KaisennHu KaisennHu commented Nov 6, 2026 •

Copy link
Copy Markdown

Summary

This change archives the release standard of vllm-project/router (vLLM Router) in the repository knowledge base as a new repository slice, knowledge/repos/vllm-router/.

The page structure follows repos/vllm-omni/ci/release-gates.md. The gate content is derived from this repository's own release pipeline and CI configuration; vLLM-Omni's CUDA L4/L5 tiers and NPU lanes are not reused.

Contents

  1. _index.md — repository entry. Records the upstream repository, the PyPI package vllm-router, the module vllm_router, aliases, and the current knowledge boundary.
  2. ci/_index.md — CI entry. Describes the four Buildkite lanes (L1 ready, L2 merge, L3 nightly, release), the pytest markers, and the GPU hardware presets.
  3. ci/release-gates.md — release standard, organised in four parts:
    1. Hard gates executed by .buildkite/release-pipeline.yml: version equality between the git tag and pyproject.toml; x86_64 and aarch64 wheel builds plus the sdist; binary and wheel smoke tests; the PyPI publication dependency chain; and the Docker release confirmation steps.
    2. Policy gates: the L2 merge result together with the latest L3 nightly run, which already covers GSM8K accuracy, the ROCm MI300 xGMI and RDMA lanes, and gRPC prefix-KV performance; unit-test coverage measured against the target set at iteration kickoff; scenario-level performance evidence without a uniform percentage threshold; requirement completion rate; remaining DI, which depends on aligning issue grading between router and agent-infer; and critical-issue ownership.
    3. Evidence rules: performance is compared per scenario for prefix-KV, L1 encode, and CPU overhead; coverage is reported against the kickoff target because CI enforces no threshold; DI is reported only after issues are graded, because the bug template carries a bug label but no severity field.
    4. Documented failure modes: version drift across the four version files; the README CI/CD section that contradicts the implemented pipeline; treating nightly builds as releases; Linux-only manylinux artifacts.

Metadata updates

  1. knowledge/repos/_index.md — adds the vLLM-Router row.
  2. doc/knowledge/SCHEMA.md — registers the vllm-router ownership tag, which the taxonomy requires before first use.

Validation

  1. python knowledge/tools/check_knowledge_tree.py — no errors and no warnings for the new slice. Three pre-existing errors remain for repos/vllm-omni/**/rules.md on a local Windows autocrlf checkout; the corresponding LF blobs measure 32678, 32700, and 32692 bytes, below 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 1529 pages.

Boundary

This change covers CI and release knowledge only. components/, models/, _routes.yaml, and an adapter are not yet established, so Direct routing does not reach this repository. The release page must not be treated as a code owner during review.

- repos/vllm-router: repo entry, CI entry, release gates (tag/pyproject version check, dual-arch wheel + smoke, PyPI and Docker publish order, policy gates)
- register in repos/_index.md and add the vllm-router ownership tag to SCHEMA.md
@KaisennHu
KaisennHu force-pushed the knowledge/vllm-router-release-standard branch from e243b04 to 9d6eb4b Compare November 6, 2026 06:35
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.

1 participant