Add deterministic vLLM-Omni release drift audit

.github/workflows/vllm-omni-release-audit.yml

@@ -0,0 +1,99 @@
+name: vLLM-Omni release drift audit
+
+on:
+  pull_request:
+    paths:
+      - .github/workflows/vllm-omni-release-audit.yml
+      - adapters/vllm_omni/manifest.yaml
+      - adapters/vllm_omni/release_baseline.yaml
+      - doc/VLLM_OMNI_RELEASE_MAINTENANCE.md
+      - knowledge/repos/vllm-omni/**
+      - tools/audit_vllm_omni_release.py
+      - tools/vllm_omni_release_audit.py
+  schedule:
+    - cron: "23 3 * * 1"
+  workflow_dispatch:
+    inputs:
+      from_ref:
+        description: Old tag or SHA; blank uses the audited baseline
+        required: false
+        type: string
+      to_ref:
+        description: New tag or SHA
+        required: true
+        default: origin/main
+        type: string
+      mode:
+        description: Scheduled checks use report-only
+        required: true
+        default: report-only
+        type: choice
+        options:
+          - report-only
+          - enforce
+
+permissions:
+  contents: read
+
+jobs:
+  audit:
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Check out vLLM-Omni
+        uses: actions/checkout@v4
+        with:
+          repository: vllm-project/vllm-omni
+          path: .audit/vllm-omni
+          fetch-depth: 0
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install audit dependency
+        run: python -m pip install PyYAML
+
+      - name: Resolve audit inputs
+        id: inputs
+        shell: bash
+        env:
+          EVENT_NAME: ${{ github.event_name }}
+          REQUESTED_FROM: ${{ inputs.from_ref }}
+          REQUESTED_TO: ${{ inputs.to_ref }}
+          REQUESTED_MODE: ${{ inputs.mode }}
+        run: |
+          read_baseline() {
+            python -c "import yaml; print(yaml.safe_load(open('adapters/vllm_omni/release_baseline.yaml'))['upstream']['$1'])"
+          }
+          if [[ "$EVENT_NAME" == "pull_request" ]]; then
+            echo "from=$(read_baseline previous_audited_sha)" >> "$GITHUB_OUTPUT"
+            echo "to=$(read_baseline audited_sha)" >> "$GITHUB_OUTPUT"
+            echo "mode=enforce" >> "$GITHUB_OUTPUT"
+          else
+            audited_sha="$(read_baseline audited_sha)"
+            echo "from=${REQUESTED_FROM:-$audited_sha}" >> "$GITHUB_OUTPUT"
+            echo "to=${REQUESTED_TO:-origin/main}" >> "$GITHUB_OUTPUT"
+            echo "mode=${REQUESTED_MODE:-report-only}" >> "$GITHUB_OUTPUT"
+          fi
+
+      - name: Run release audit
+        run: |
+          python tools/audit_vllm_omni_release.py \
+            --from "${{ steps.inputs.outputs.from }}" \
+            --to "${{ steps.inputs.outputs.to }}" \
+            --repo .audit/vllm-omni \
+            --mode "${{ steps.inputs.outputs.mode }}" \
+            --json-output "${RUNNER_TEMP}/vllm-omni-release-audit.json"
+
+      - name: Upload machine report
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: vllm-omni-release-audit
+          path: ${{ runner.temp }}/vllm-omni-release-audit.json
+          if-no-files-found: warn

README.md

@@ -123,5 +123,6 @@ Strict 模式需要用户明确提出。具体用法见
 
 - [安装和 MCP 配置](doc/MCP.md)
 - [知识库贡献规范](knowledge/CONTRIBUTING.md)
+- [vLLM-Omni 发版漂移审计](doc/VLLM_OMNI_RELEASE_MAINTENANCE.md)
 - [项目设计与实现](doc/)
 - [评测说明](eval/README.md)

adapters/vllm_omni/release_baseline.yaml

@@ -0,0 +1,144 @@
+schema_version: 1
+
+upstream:
+  repository: vllm-project/vllm-omni
+  previous_audited_sha: 5d44868e918ecf9d3a6f1158c45acfad5989e1a7
+  audited_ref: v0.26.0rc1
+  audited_sha: 807db6efd70ff2e9b55a63d6e1b0530e2b74f8f2
+
+# These fingerprints are generated from sorted registry/deploy inventories.
+# They make the baseline exact without duplicating hundreds of entries here.
+inventories:
+  autoregressive:
+    count: 72
+    sha256: 79afac742125003b8d65a02e0e6b53398c4fc5f1c6e765fa8e1eb31f74871195
+  diffusion:
+    count: 61
+    sha256: f5c0b782cb98f20a9c7c77b696fda620a12d0899fd9e8abc9ccb696fe0c457e1
+  pipelines:
+    count: 46
+    sha256: ce4f6edb8748e030f04bf7a6fe038231453cd1cd75f64311c898acff8c14b14b
+  deploy_yamls:
+    count: 71
+    sha256: fb4c591580262a2dc8d03dd8c7fd681d98486f31007696d56adf349c24c25147
+
+# Structural review owners for changed upstream paths. This is deliberately
+# separate from manifest.modules: changing that runtime map alters rebase fan-out.
+path_owners:
+  benchmark:
+  - benchmarks/
+  - vllm_omni/benchmarks/
+  ci:
+  - .buildkite/
+  - .github/
+  - docker/
+  configuration:
+  - vllm_omni/config/
+  - vllm_omni/deploy/
+  diffusion:
+  - vllm_omni/diffusion/
+  distributed:
+  - vllm_omni/distributed/
+  documentation:
+  - .claude/skills/
+  - docs/
+  - examples/
+  - recipes/
+  - realtime_video_prompt_interaction_protocol.md
+  model-executor:
+  - vllm_omni/attention/
+  - vllm_omni/experimental/
+  - vllm_omni/inputs/
+  - vllm_omni/lora/
+  - vllm_omni/model_executor/
+  - vllm_omni/model_extras/
+  - vllm_omni/outputs/
+  - vllm_omni/platforms/
+  - vllm_omni/plugins/
+  - vllm_omni/quantization/
+  - vllm_omni/reasoning/
+  - vllm_omni/sample/
+  - vllm_omni/tokenizers/
+  - vllm_omni/transformers_utils/
+  - vllm_omni/utils/
+  - vllm_omni/worker/
+  observability:
+  - vllm_omni/metrics/
+  - vllm_omni/profiler/
+  scheduler:
+  - vllm_omni/core/
+  serving:
+  - vllm_omni/engine/
+  - vllm_omni/entrypoints/
+  testing:
+  - tests/
+  tooling:
+  - tools/
+  runtime-core:
+  - vllm_omni/__init__.py
+  - vllm_omni/data_entry_keys.py
+  - vllm_omni/request.py
+
+# Direct review destination for every structural owner above. The audit checks
+# that each route has a real current knowledge entrypoint.
+owner_documents:
+  benchmark:
+  - knowledge/repos/vllm-omni/benchmark/_index.md
+  ci:
+  - knowledge/repos/vllm-omni/ci/_index.md
+  configuration:
+  - knowledge/repos/vllm-omni/components/configuration/rules.md
+  diffusion:
+  - knowledge/repos/vllm-omni/components/diffusion/rules.md
+  distributed:
+  - knowledge/repos/vllm-omni/components/distributed/_index.md
+  documentation:
+  - knowledge/general/docs/_index.md
+  model-executor:
+  - knowledge/repos/vllm-omni/components/model-executor/rules.md
+  observability:
+  - knowledge/repos/vllm-omni/rules.md
+  scheduler:
+  - knowledge/repos/vllm-omni/components/scheduler/rules.md
+  serving:
+  - knowledge/repos/vllm-omni/components/serving/rules.md
+  testing:
+  - knowledge/repos/vllm-omni/ci/_index.md
+  tooling:
+  - knowledge/repos/vllm-omni/rules.md
+  runtime-core:
+  - knowledge/repos/vllm-omni/components/model-executor/rules.md
+
+ignored_paths:
+- pattern: .gitignore
+  reason: repository metadata
+- pattern: .pre-commit-config.yaml
+  reason: repository tooling
+- pattern: .readthedocs.yaml
+  reason: documentation tooling
+- pattern: CODE_OF_CONDUCT.md
+  reason: community metadata
+- pattern: CONTRIBUTING.md
+  reason: community metadata
+- pattern: LICENSE
+  reason: legal metadata
+- pattern: README.md
+  reason: repository overview
+- pattern: SECURITY.md
+  reason: community metadata
+- pattern: setup.py
+  reason: package metadata
+- pattern: pyproject.toml
+  reason: package metadata
+- pattern: requirements/
+  reason: package dependencies
+
+pin_documents:
+- doc/KNOWLEDGE.md
+- knowledge/repos/vllm-omni/models/catalog.md
+- knowledge/repos/vllm-omni/components/configuration/_index.md
+- knowledge/repos/vllm-omni/components/diffusion/_index.md
+- knowledge/repos/vllm-omni/components/distributed/_index.md
+- knowledge/repos/vllm-omni/components/model-executor/_index.md
+- knowledge/repos/vllm-omni/components/scheduler/_index.md
+- knowledge/repos/vllm-omni/components/serving/_index.md

doc/KNOWLEDGE.md

@@ -42,7 +42,8 @@ extensions merely because they are absent from the common source.
   history was removed — used for page frontmatter).
 - **Code-mirror pin:** the `knowledge/repos/vllm-omni/components/` source maps
   are verified against vllm-omni `main @
-  238fc0a609311235a671940cf209a7eb72c1dc29` (every stated path exists there).
+  807db6efd70ff2e9b55a63d6e1b0530e2b74f8f2`. The canonical machine baseline
+  is `adapters/vllm_omni/release_baseline.yaml`.
 
 ## Layout: general vs repo-specific
 The knowledge base separates the two, and so do we — the whole tree is **shared**,

doc/VLLM_OMNI_RELEASE_MAINTENANCE.md

@@ -0,0 +1,51 @@
+# vLLM-Omni 发版漂移审计
+
+InferMatrixCopilot 的模型清单、代码 owner 路由和源码引用会随着
+vLLM-Omni 发版而过期。这个审计只读取两个 Git 提交和本仓库声明，不会 checkout
+目标版本，也不会自动修改知识规则。
+
+## 本地运行
+
+先在 vLLM-Omni checkout 中 fetch 需要比较的 tag 或 SHA，然后运行：
+
+```powershell
+python tools/audit_vllm_omni_release.py `
+  --from 5d44868e `
+  --to v0.26.0rc1 `
+  --repo D:\path\to\vllm-omni `
+  --json-output $env:TEMP\vllm-omni-release-audit.json
+```
+
+审计覆盖：
+
+- AR、Diffusion 和 pipeline registry；
+- deploy YAML；
+- 新增、修改、删除和重命名路径；
+- changed path 到 Direct 知识 owner 的机器路由；
+- adapter runtime module 未覆盖的 changed path；
+- active knowledge 中受删除或重命名影响的 `sources:`；
+- component、model catalog 与 canonical baseline 的 source pin。
+
+同样的提交和 baseline 会产生等价 JSON；报告不包含时间戳和本机 checkout 路径。
+默认 `--mode enforce`：存在未解释漂移时退出 1，输入或 Git 失败时退出 2。
+`--mode report-only` 仍报告 `DRIFT`，但退出 0，供定时巡检使用。
+
+## 更新一个 release
+
+1. 用当前 `release_baseline.yaml` 的 audited SHA 作为 `--from`，新 tag 作为
+   `--to`，先跑 `report-only`。
+2. 人工确认 registry、deploy、路径 owner 和知识来源变化。
+3. 把旧的 `audited_sha` 移到 `previous_audited_sha`，再只更新报告证明已经漂移的
+   baseline、catalog、source map 或 manifest；不要自动生成语义规则。
+4. 用同一组 `--from/--to` 跑 `enforce`，再运行知识 validator 和相关 pytest。
+5. JSON 是临时证据，不提交到 `knowledge/`；完成后删除。
+
+## PR 学习与 release 审计是两件事
+
+合并 PR 的复盘只提炼可复用、可执行的 owner 规则，且必须等最终修复、review thread
+和 CI 结果稳定后再做。Release 审计只对账结构事实，不把 PR 内容、事故过程或审计报告
+写进知识树，也不因为上游发版自动升级 InferMatrixCopilot 版本或依赖。
+
+CI 每周对 upstream `main` 运行 `report-only`。修改 baseline 的 PR 会自动用
+`previous_audited_sha → audited_sha` 运行 `enforce`；Actions 手动运行可用
+`report-only` 检查任意区间，`enforce` 只接受当前 baseline 声明的升级区间。

knowledge/repos/vllm-omni/ci/guides/test-tiers.md

@@ -1,16 +1,16 @@
 ---
 title: "测试分级（L1–L5）与 pytest markers"
 created: 2026-07-16
-updated: 2026-07-16
+updated: 2026-07-31
 type: guide
 tags: [vllm-omni, ci]
-sources: [docs/contributing/ci/CI_5levels.md, docs/contributing/ci/tests_markers.md]
+sources: [docs/contributing/ci/test_system_overview.md, docs/contributing/ci/test_writing_guide.md]
 ---
 
 # 测试分级（L1–L5）与 pytest markers
 
-官方 spec：`docs/contributing/ci/CI_5levels.md` + `tests_markers.md`
-（`main @ 5c390096` 复核）。测试金字塔五级 + Common 规范
+官方 spec：`docs/contributing/ci/test_system_overview.md` +
+`test_writing_guide.md`（`main @ 807db6ef` 复核）。测试金字塔五级 + Common 规范
 （PR 模板/checklist 与 CI 失败说明）。
 
 ## 五级定义

knowledge/repos/vllm-omni/components/configuration/_index.md

@@ -14,6 +14,8 @@ sources: ["claude-workflow-starter-private@296ea45", vllm_omni/config/]
 - 主要测试：`tests/config/`、`tests/test_config_factory.py`、`tests/test_diffusion_config_propagation.py`，以及各公开入口附近的配置测试
 - 部署配置：`vllm_omni/deploy/*.yaml`，以及 `pipeline_registry.py`、
   `endpoint_policy.py`、`server_settings.py`、`yaml_util.py`、`composable_parallel/`
+- 源码校验：以上路径在 `main @ 807db6ef` 验证存在；机器基线见
+  `adapters/vllm_omni/release_baseline.yaml`
 
 ## 什么时候查这里
 

knowledge/repos/vllm-omni/components/diffusion/_index.md

@@ -10,7 +10,7 @@ sources: []
 # Diffusion
 
 - 源码入口：`vllm_omni/diffusion/` 全树，含 16 个子模块：attention、cache、distributed、executor、hooks、layers、lora、model_loader、models、offloader、postprocess、profiler、quantization、sched、utils、worker
-- 源码校验：以上子模块均已在 `main @ 238fc0a6`（此前亦在 `dev/vllm-align @ 4f2b32c` 验证，结果一致） 验证存在
+- 源码校验：以上子模块均已在 `main @ 807db6ef` 验证存在
 - 主要职责：多个 diffusion 模型共用的 pipeline、执行循环、scheduler 接入和运行机制
 
 ## 什么时候查这里

knowledge/repos/vllm-omni/components/distributed/_index.md

@@ -14,7 +14,7 @@ sources: [vllm_omni/distributed/omni_connectors/, vllm_omni/distributed/omni_coo
   `vllm_omni/distributed/omni_coordinator/`（协调器与 load balancer）
 - 知识面另覆盖跨 stage ZMQ 路由/端口分配（`vllm_omni/engine/stage_engine_startup.py::OmniMasterServer`）
   ——组件划分服务知识归属，与 manifest 运行时粒度不同
-- 源码校验：以上路径与下列锚点均已在 `main @ 5c390096` 验证存在：
+- 源码校验：以上路径与下列锚点均已在 `main @ 807db6ef` 验证存在：
   `OmniConnectorBase`（connectors/base.py:12）、`OmniKVTransferManager`
   （kv_transfer_manager.py:341）、`LoadBalancer` 三实现（load_balancer.py:39/64/74/102）、
   `OmniMasterServer._allocate_route_locked`（stage_engine_startup.py:254）

knowledge/repos/vllm-omni/components/model-executor/_index.md

@@ -9,8 +9,9 @@ sources: []
 
 # Model Executor
 
-- 源码入口：`vllm_omni/model_executor/`（layers、model_loader、models、stage_configs、stage_input_processors）、`vllm_omni/worker/`（gpu_*_worker、gpu_*_model_runner、mixins）、`vllm_omni/inputs/`（runner 输入预处理：data.py、preprocess.py）和设备平台层 `vllm_omni/platforms/<cuda|musa|npu|rocm|xpu>/platform.py`
-- 源码校验：以上路径均已在 `main @ 238fc0a6`（此前亦在 `dev/vllm-align @ 4f2b32c` 验证，结果一致） 验证存在；旧的 `platforms/*/worker/` 布局在该提交已不存在（平台目录只含 `platform.py`）
+- 源码入口：`vllm_omni/model_executor/`（layers、model_loader、models、stage_input_processors）、`vllm_omni/worker/`（gpu_*_worker、gpu_*_model_runner、mixins）、`vllm_omni/inputs/`（runner 输入预处理：data.py、preprocess.py）和设备平台层 `vllm_omni/platforms/<cuda|musa|npu|rocm|xpu>/`
+- 源码校验：以上路径均已在 `main @ 807db6ef` 验证存在；stage 配置已经迁移到
+  `vllm_omni/deploy/`，NPU/XPU 等平台可继续拥有自己的 worker 覆盖
 - 测试入口：共享 runner 行为看 `tests/worker/`，具体模型 consumer 看 `tests/model_executor/`
 - 主要职责：AR/LLM stage、stage 配置、并行与设备启动、runner 到模型的输入预处理合同和跨阶段数据桥接
 

knowledge/repos/vllm-omni/components/scheduler/_index.md

@@ -11,7 +11,7 @@ sources: [vllm_omni/core/sched/omni_ar_scheduler.py, vllm_omni/core/prefix_cache
 
 - 源码入口：`vllm_omni/core/sched/`（`omni_ar_scheduler.py`、`omni_generation_scheduler.py`、
   `omni_scheduler_mixin.py`、`omni_scheduling_coordinator.py`）和 `vllm_omni/core/prefix_cache.py`
-- 源码校验：以上路径与下列类均已在 `main @ 5c390096` 验证存在：`OmniARScheduler`（:50）、
+- 源码校验：以上路径与下列类均已在 `main @ 807db6ef` 验证存在：`OmniARScheduler`（:50）、
   `OmniARAsyncScheduler`（:928）、`KVCacheTransferData`（:40）、`OmniGenerationScheduler`（:42）、
   `OmniSchedulerMixin`（:40）、`OmniTensorPrefixCache`（prefix_cache.py:33）
 - 官方设计文档：`docs/design/module/ar_module.md`（继承关系、请求流转图）

knowledge/repos/vllm-omni/components/serving/_index.md

@@ -10,7 +10,7 @@ sources: []
 # Serving
 
 - 主要源码入口：`vllm_omni/entrypoints/`（cli、openai、openpi 及 omni/async_omni 入口）和 `vllm_omni/engine/`（orchestrator、stage engine core、stage pool/runtime、output processor）
-- 源码校验：以上路径均已在 `main @ 238fc0a6`（此前亦在 `dev/vllm-align @ 4f2b32c` 验证，结果一致） 验证存在
+- 源码校验：以上路径均已在 `main @ 807db6ef` 验证存在
 - 主要职责：用户入口、请求解析、在线服务和 engine 边界
 
 ## 什么时候查这里