[VideoGenX] run matched Wan workloads on SGLang and vLLM-Omni

.github/workflows/test-h3-video.yml

@@ -41,7 +41,6 @@ concurrency:
 jobs:
   test:
     name: H3 CPU harness and workflow checks
-    if: ${{ github.event.pull_request.draft != true }}
     runs-on: ubuntu-latest
     timeout-minutes: 15
     steps:
@@ -58,7 +57,9 @@ jobs:
           check_env_vars PYTHONPATH
           uv run --no-project --python 3.12 \
           --with 'av==16.1.0' --with 'numpy==2.3.5' \
-          --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' python -m pytest -q
+          --with 'pyyaml==6.0.3' \
+          --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' \
+          --with 'rfc3339-validator==0.1.4' python -m pytest -q
       - name: Check entry shell syntax
         run: |
           bash -n experimental/video-generation/runtime-entry.example.sh
@@ -76,7 +77,6 @@ jobs:
 
   staged-power-python310:
     name: H3 staged power on Python 3.10
-    if: ${{ github.event.pull_request.draft != true }}
     runs-on: ubuntu-latest
     timeout-minutes: 5
     steps:

experimental/video-generation/README.md

@@ -2,6 +2,71 @@
 
 **English** | [中文](README_zh.md)
 
+## Wan2.2 local runner integration
+
+The client accepts `Wan-AI/Wan2.2-T2V-A14B-Diffusers` on SGLang and vLLM-Omni;
+the supervised `serving-smoke` path starts SGLang. A pinned eight-H100
+vLLM-Omni server completed one Wan request. The full workload, matched performance,
+quality, and production delivery remain unverified.
+
+The [Wan plan](mvp/wan22-serving.plan.json) freezes checkpoint
+`5be7df9619b54f4e2667b2755bc6a756675b5cd7`, four video-only prompts, one warmup,
+832×480, 81 frames at 16 FPS (5.0625 seconds), 40 steps, high/low expert guidance
+4/3, flow shift 12, and an explicitly empty negative prompt. This is a new workload,
+not a matched H3 comparison or the upstream default negative prompt. The protocol
+was inspected at SGLang `71de97b264b04dcd514cf904003028aefe9775c8`. That runtime
+ignores per-request Wan flow-shift overrides, so only 12 is admitted; high-stage
+guidance must exceed 1 for the runtime to apply both expert guidance values.
+The [shared Wan spec](specs/wan.yaml) retains that four-case plan as `frozen-3733`
+and defines `shared-v1` with two seeds per prompt. The [serving note](specs/wan-vllm-omni-serving.md)
+records the pinned vLLM-Omni setup and single-request evidence. The spec loader
+is shared with H3 and keeps the plan hash separate from source citations.
+
+Preview requests without a server or GPU, from `experimental/video-generation`:
+
+```bash
+python -m evaluator.cli run mvp/wan22-serving.plan.json --runtime sglang
+```
+
+The [GPU template](mvp/wan22-gpu-job.template.json) retains unapproved authorization,
+unresolved paths and an empty file manifest. Before separately authorized execution,
+bind it to task-owned SemiAnalysis resources and the compatible prepared runtime,
+verify all checkpoint files, and set the actual budget. The local supervised entry is
+`python -m evaluator.cli serving-smoke --spec <prepared-spec.json> --output <new-evidence-dir> --concurrencies 1`.
+Its first cell has four measured requests plus one warmup. The template does not
+establish one-GPU memory fit.
+
+Initial support is NVIDIA, TP1 and pure Ulysses over 1/2/4/8 GPUs, with explicit
+DiT/text-encoder/VAE offload booleans and layerwise offload disabled. The launcher
+disables runtime warmup, CFG parallelism and Ring, and uses a job-owned persistent
+output directory so the polled `/v1/videos/{id}/content` route can deliver the MP4.
+The harness records its own warmup. H3-only launch flags/timing patches, AMD,
+and paired quality qualification are rejected for this Wan launcher. The runner
+also accepts an operator-managed vLLM-Omni endpoint using the audited multipart transport.
+The policy records only an uncalibrated identity; it contains no audio thresholds.
+
+Media validation requires exact geometry, cadence, frame count and no audio stream.
+Audio generation and quality qualification are unsupported; no audio/A-V score is
+invented. Missing quality and quality/SLO goodput stay null in the dashboard.
+Timeouts retain unknown remote completion and stop further submissions; deleting a
+job in this pinned server does not prove cancellation.
+
+The local supervised entry writes a `video_serving_smoke_matrix` and retains failed
+and unstarted request counts. The operator-managed `run --execute` entry writes a
+separate run receipt. The parent H3 CI now launches fresh jobs through native
+srt-slurm video matrix rows. Wan has no native recipe or config row, so these local
+entries do not submit or qualify a Wan CI job. A native Wan row needs its own
+reviewed recipe and artifact wiring. The paired H3 `result.json` exporter is not
+used for Wan. The app accepts the canonical Diffusers ID and the earlier short
+display ID, and rejects conflicting model declarations.
+
+CPU checks cover both HTTP transports with encoded fixtures, full video-only
+decoding, launch previews, failure seals and dashboard parsing. The retained H100
+record proves one request completed; it does not qualify the shared workload or a
+production path. Fixtures are never included as measured results.
+
+## Existing H3 lane
+
 This experimental lane runs the existing H3 supervisor inside InferenceX CI on
 prepared SemiAnalysis NVIDIA resources. Its first target is a bounded same-build smoke:
 original generated MP4s, full video/audio validation, measured requests, and
@@ -250,7 +315,8 @@ Regression mode additionally requires the existing calibrated acceptance gate.
 cd experimental/video-generation
 PYTHONPATH=../../inferencex-e2e uv run --no-project --python 3.12 \
   --with 'av==16.1.0' --with 'numpy==2.3.5' \
-  --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' python -m pytest -q
+  --with 'pyyaml==6.0.3' --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' \
+  --with 'rfc3339-validator==0.1.4' python -m pytest -q
 bash -n runtime-entry.example.sh
 ```
 

experimental/video-generation/README_zh.md

@@ -2,6 +2,61 @@
 
 [English](README.md) | **中文**
 
+## Wan2.2 本地 runner 接入
+
+现有 client 可通过 SGLang 或 vLLM-Omni 运行
+`Wan-AI/Wan2.2-T2V-A14B-Diffusers`；受控 `serving-smoke` 路径启动 SGLang。
+一台固定版本的八卡 H100 vLLM-Omni
+服务完成了一次 Wan 请求。完整工作负载、匹配的性能对比、质量及生产交付仍未验证。
+
+[Wan plan](mvp/wan22-serving.plan.json) 固定 checkpoint
+`5be7df9619b54f4e2667b2755bc6a756675b5cd7`，四条纯视频提示词、一次独立 warmup，
+832×480、81 帧、16 FPS（5.0625 秒）、40 步、两阶段 guidance 4/3、flow shift 12，
+并显式使用空 negative prompt。这是新工作负载，不是与 H3 匹配的对照，也不是上游
+默认 negative prompt。协议按 SGLang `71de97b264b04dcd514cf904003028aefe9775c8`
+核对。该版本忽略请求级 Wan flow-shift 覆盖，因此仅接受 12；第一阶段 guidance 必须
+大于 1，运行时才会应用两个 expert 的 guidance 值。
+[共享 Wan spec](specs/wan.yaml) 用 `frozen-3733` 保留原来的四个 case，并在
+`shared-v1` 中为每条提示词加入第二个 seed。[服务记录](specs/wan-vllm-omni-serving.md)
+保留固定版本的 vLLM-Omni 配置和单次请求证据。spec loader 与 H3 共用，plan hash
+不包含来源引用。
+
+在 `experimental/video-generation` 下无 GPU、无服务地预览请求：
+
+```bash
+python -m evaluator.cli run mvp/wan22-serving.plan.json --runtime sglang
+```
+
+[GPU 模板](mvp/wan22-gpu-job.template.json) 保持未批准授权、未填写路径和空文件清单。
+另行授权实际运行前，需要绑定任务拥有的 SemiAnalysis 资源及兼容的已有 runtime，
+核实全部 checkpoint 文件，并设置实际预算。本地受控入口是
+`python -m evaluator.cli serving-smoke --spec <prepared-spec.json> --output <new-evidence-dir> --concurrencies 1`。
+第一格为四条测量请求加一次 warmup；模板不能证明单卡显存足够。
+
+初始支持 NVIDIA、TP1、1/2/4/8 卡纯 Ulysses，显式设置 DiT/text-encoder/VAE
+offload，禁用 layerwise offload。launcher 关闭运行时 warmup、CFG 并行和 Ring，
+使用本次任务拥有的持久输出目录，确保轮询后的 `/v1/videos/{id}/content` 能下载 MP4。
+harness 单独记录 warmup。该 Wan launcher 拒绝 H3 专用启动参数/计时补丁、AMD
+及配对质量准入。runner 也可通过已核对的 multipart 协议连接由操作员管理的
+vLLM-Omni 服务。policy 仅记录未校准身份，不含音频阈值。
+
+媒体验证要求分辨率、时间间隔、帧数符合计划，且无音频流。音频生成和质量准入为
+unsupported，不生成音频或 A/V 分数。Dashboard 缺失质量及质量/SLO goodput 仍为 null。
+超时后保留远端是否完成的未知状态并停止后续提交；该版本删除 job 不代表真正取消。
+
+本地受控入口写入 `video_serving_smoke_matrix`，保留失败及未启动的请求数。
+由操作员管理的 `run --execute` 入口另写一份运行收据。父分支的 H3 CI 现在通过原生
+srt-slurm video matrix 行启动新作业。Wan 尚无对应的原生 recipe 或 config 行，
+因此这些本地入口不会提交或验收 Wan CI 作业。原生 Wan 行仍需经审核的 recipe
+及产物接线。Wan 不使用 H3 配对 `result.json` exporter。App 接受规范 Diffusers ID
+与此前短显示 ID，并拒绝冲突的模型声明。
+
+CPU 检查覆盖两种使用编码 fixture 的 HTTP transport、纯视频完整解码、启动预览、
+失败封存及 Dashboard 解析。保留的 H100 记录只证明一次请求已完成，不能代表共享
+工作负载或生产路径通过验收。fixture 不进入实测结果。
+
+## 既有 H3 路径
+
 此实验任务在 InferenceX CI 内使用 SemiAnalysis H200 资源运行现有 H3
 supervisor。首个目标是有时间上限的同版本冒烟测试：保留实际生成的 MP4，
 完整验证视频和音频，记录请求耗时，并验证清理结果。此任务不写入原生
@@ -135,9 +190,9 @@ H200 硬件盘点、CPU 导出）与 `h3-fidelity.yml`（保留媒体比较）
 
 ## 结果与本地检查
 
-成功执行还会发布[前端结果契约](RESULTS_zh.md)：版本化的 `result.json`、
+成功执行还会发布[前端结果契约（英文）](RESULTS.md)：版本化的 `result.json`、
 逐卡功率序列、分阶段能量与覆盖率，以及可离线打开的 `power-report.html`。
-`h3-results-<run-id>-<attempt>` 包含索引、JSON Schema、双语指标说明，以及每次
+`h3-results-<run-id>-<attempt>` 包含索引、JSON Schema、指标说明，以及每次
 原始执行的媒体、测量、日志和报告。原始执行身份与校验文件和本次导出身份分开保留。
 遥测无效时功率和能量为空；导出失败会保留错误日志并返回失败状态。
 
@@ -170,7 +225,7 @@ gh workflow run h3-video.yml --repo SemiAnalysisAI/InferenceX \
 配置、实际执行、默认及最大功率上限均为 700 W；历史生成时的设置仍未知。
 原始盘点记录中的 unknown 分类保持不变；导出器使用新 producer 提交解析原始
 XML，同时支持带或不带 `0x` 的 PCI 编号，不重新查询硬件。
-[已保留的 A/A 测量及限制](RESULTS_zh.md#已观测的-aa-证据)提供具体数据。
+[已保留的 A/A 测量及限制（英文）](RESULTS.md#observed-aa-evidence)提供具体数据。
 
 原始工作负载失败时仍上传失败证据，但不会启动成功结果导出。
 
@@ -189,7 +244,8 @@ serving 运行由其基准测试作业上传 `h3-results-<run-id>-<attempt>`；
 cd experimental/video-generation
 PYTHONPATH=../../inferencex-e2e uv run --no-project --python 3.12 \
   --with 'av==16.1.0' --with 'numpy==2.3.5' \
-  --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' python -m pytest -q
+  --with 'pyyaml==6.0.3' --with 'pytest>=8,<9' --with 'jsonschema>=4,<5' \
+  --with 'rfc3339-validator==0.1.4' python -m pytest -q
 bash -n runtime-entry.example.sh
 ```
 

experimental/video-generation/RESULTS.md

@@ -1,7 +1,5 @@
 # H3 backend result contract
 
-**English** | [中文](./RESULTS_zh.md)
-
 `result.json` is the frontend entry point for a downloaded H3 CI artifact. Its
 `schema_version` is **1.0.0**; [result.schema.json](./result.schema.json) describes
 its external shape. Reject unknown versions. This contract does not ingest a
@@ -200,3 +198,83 @@ Each duration is a separate frozen workload. Do not pool these durations or
 interpret their latency difference as a regression. One measured clip per role
 provides point estimates only; workload execution passed while regression
 remains uncalibrated and inconclusive.
+
+## Offline deployment-frontier replay
+
+`evaluator.mvp_deployment_frontier` selects retained, normalized dashboard
+observations for a local comparison; it does not collect new measurements or
+qualify an artifact. This is the backend counterpart of the video-only wrapper in
+[App #1193](https://github.com/SemiAnalysisAI/InferenceX-app/pull/1193), delivered
+with [Backend #2916](https://github.com/SemiAnalysisAI/InferenceX/pull/2916).
+The shared LLM/calculator Pareto implementation is unchanged.
+
+Pass a JSON object with `points`. Each point supplies `id`, finite numeric `x` and
+`y`, `hardwareKey`, canonical `cohortKey` (workload plus generation settings),
+canonical `deploymentKey` (including replicas, parallel layout, offload, batching
+and engine/runtime), and the producer's `queueingStatus` (`unqueued`, `queueing`,
+or `unknown`). Optional `hardwareHealth.status: "fail"` excludes a point even
+when dominated deployments are requested. Missing canonical cohorts remain
+isolated points. Unknown request capacity and missing axes cannot rank. Exact
+coordinate ties retain distinct measured identities; a repeated observation of
+one deployment does not create a connecting deployment line. No interpolation
+or extrapolation is performed.
+
+With `--quality-required`, the reader checks the source-bound `quality` assessment;
+an opaque `qualifiedQualityCohort` string cannot grant admission. The assessment
+must declare `scale: "ordinal_0_to_4"` and preserve all seven original rubric IDs:
+`prompt_adherence`, `visual_fidelity`, `temporal_consistency`, `motion_plausibility`,
+`audio_quality`, `audio_content`, and `av_sync`. Zero is a real score, not missing
+information. Legacy 1–5 scores or renamed dimensions are not silently converted.
+The registry remains `human.absolute_dimension_rating`, version `0.2.0-draft`.
+
+Every critical dimension must have a passing decision, complete target-clip
+coverage, identified evaluator, and a calibrated frozen rule with provenance.
+`--quality-metric` selects the displayed dimension (default `prompt_adherence`);
+`--quality-threshold` can only tighten that dimension's frozen threshold. Other
+dimensions still require their own frozen rules. The derived quality cohort includes
+all seven evaluator and calibration identities, not only the selected dimension.
+Different workload/generation or quality cohorts never dominate each other or share
+a connecting line. These checks validate recorded contract fields; they do not
+establish that human judgments or calibration are scientifically valid, or turn
+media integrity/fidelity into perceptual quality.
+
+```bash
+cd experimental/video-generation
+python -m evaluator.mvp_deployment_frontier normalized-observations.json \
+  --output frontier.json --x-better lower --y-better higher
+```
+
+The JSON output retains input identities, selected observations, each separate
+frontier and exclusion reasons. `--include-dominated` matches the chart's Optimal
+Only off state; `--quality-required` applies the qualification filter first.
+`tests/fixtures/deployment-frontier.json` contains shared synthetic direction/tie
+cases for the backend and App tests, never measured benchmark evidence.
+
+## Optional source-bound dashboard observations
+
+[`dashboard-observations.schema.json`](./dashboard-observations.schema.json)
+defines an additive `dashboard-observations.json` sidecar, independent of
+`result.json` version 1.0.0. Its root contains `schemaVersion: 1`, `sourceRunId`,
+`sourceSha`, and a nonempty `cells` map. Each known cell binds `runSha256` and
+`specSha256` to that cell's existing measured run and specification; optional
+`deployment`, `hardwareHealth`, and `quality` objects preserve unknowns as null.
+The sidecar's SHA256 must appear in both `SHA256SUMS` and
+`manifest.evidence["dashboard-observations.json"]`. Ingestion verifies the raw
+bytes, source run/commit, cell existence and run/spec hashes before exposing any
+of its fields. Schema validation alone cannot establish those relationships.
+
+Deployment fields distinguish observed `batchSize` from configured
+`maxBatchSize`, and retain Ring/CFG, individual offload flags, precision scope,
+engine and generation identity. A `hardwareHealth` record carries its status,
+reason and evidence; incomplete monitoring remains unknown. Quality stores
+the required `ordinal_0_to_4` scale and separate original rubric dimensions with
+contract/rubric/evaluator identities and hashes, sample coverage and a frozen
+calibration record. Nullable unjudged data is valid
+contract data, but cannot qualify a frontier. A syntactically valid `pass` is
+also insufficient without all seven calibrated protocol checks in the App and
+backend frontier reader.
+
+The existing exporter does not create this optional sidecar automatically.
+Original archives and source seals remain immutable. Explicit local curator
+enrichment can exercise these fields for a preview, but must be labelled as
+unsealed replay data; it is not equivalent to this sealed ingestion contract.