[edwingao28]

What changed

An unofficial VideoGenX run for Wan2.2 T2V A14B (Wan2.2 B200 vLLM-Omni benchmark, run 43653414377) rendered an empty chart with "No matching stored observation for this run", both from a local replay and from the live artifact read in production. Root cause: tradeoff.ts (workloadInfo) and results.ts (resultMetadata) only accepted a workload whose plan carries generation.duration_seconds. The Wan spec requests frame_count: 81 and fps: 16 with no nominal duration, so its points had no workload key and were dropped. A shared generationDuration helper in bundle.ts now returns duration_seconds when present and frame_count / fps otherwise. H3 plans keep their nominal duration, and plans with neither stay incomplete.

Also:

• ?unofficialRun= (the spelling used by the PR visualizer comment and accepted by the inference provider) was ignored by /vgx, which only read lowercase unofficialrun(s). parseResultsState now matches the key in any case, and the URL commit strips every case variant.
• b300 joins RESULTS_HARDWARE, with a VideoGenX colour (light #6d28d9, dark #a78bfa, both above 3:1 on their cards) and its GPU spec name, so the Wan2.2 B300 row plots once it runs.
• VIDEO_HISTORY_PROJECTION goes from 4 to 5 because projected entries for frame-count plans change, and the source digest is re-pinned.
• API docs reviewed for the changed sources: the r_duration description, the dashboard-views skill reference and videogenx-dashboard.md now describe the frame-count duration and the case-insensitive overlay key. The shared-source digests and the skill integrity manifest are refreshed.

Verification

• Unit tests: bundle.test.ts (helper cases), history.test.ts (a Wan-shaped plan gets a workload key, durationSeconds = 81/16, and stays unkeyed without frames), results.test.ts (unofficialRun= and UnofficialRuns= parse). Before the fix all three failed, e.g. expected null to be '43653414377'. All pass after it: vitest run src/components/video-benchmark src/lib/video gives 412 passed.
• Component / browser tests: none added. Checked by rendering /vgx?unofficialRun=43653414377 against a replay of the real run artifact.
• Widths inspected: 393px and 1280px.
• Checks run locally: full app vitest (12194 passed), tsc --noEmit, oxlint, oxfmt, check:typography. CI is green.
• Runtime check (Next.js): next dev with the replay manifest compiles with no runtime errors. The B200 (vLLM-Omni) point plots at about 16.9 s median latency, and the header reads Wan-AI/Wan2.2-T2V-A14B-Diffusers · 832×480 · 5.0625 s · 16 fps · 40 steps · vLLM-Omni.

Screenshots

Before: the empty chart with "No matching stored observation for this run". After: the B200 point and legend entry for run 43653414377. Both captured at 1280px and 393px and shared in review.

Not verified

• Production /vgx after deploy (it depends on the live artifact read and GITHUB_TOKEN).
• The history backfill rebuilding v5 snapshots in production.
• B300 rendering with real data (no B300 run exists yet).
• The zh locale for the new duration label. The clip length displays as an exact value (5.0625 s).
• The visualizer comment on InferenceX PRs does not link /vgx yet. That change belongs in the InferenceX repo.

中文

改动

Wan2.2 T2V A14B 的非官方 VideoGenX 运行（Wan2.2 B200 vLLM-Omni 基准，运行 43653414377）显示空图表和“No matching stored observation for this run”，本地回放与生产环境的实时产物读取都是这样。根因：tradeoff.ts 与 results.ts 只接受带 generation.duration_seconds 的计划。Wan 规格只给出 frame_count: 81 和 fps: 16，没有标称时长，因此数据点没有负载键，被过滤掉。bundle.ts 中新增共享的 generationDuration：有 duration_seconds 时直接使用，否则用 frame_count / fps。H3 计划仍使用标称时长，两者都没有的计划仍视为不完整。

此外：

• PR 可视化评论使用 ?unofficialRun=，推理看板也接受这个写法，但 /vgx 只认小写 unofficialrun(s)，因此会忽略它。现在 parseResultsState 不区分大小写，提交 URL 时会移除所有大小写变体。
• b300 加入 RESULTS_HARDWARE，并配上 VideoGenX 配色和 GPU 规格名，Wan2.2 B300 配置 运行后即可显示。
• 帧数计划的投影条目会变化，因此 VIDEO_HISTORY_PROJECTION 由 4 升到 5，并重新固定源码摘要。

验证

• 新增测试在修复前全部失败，修复后 412 个测试通过。tsc、oxlint、oxfmt、typography 检查均通过。
• 在 393px 和 1280px 宽度下，用真实产物回放检查 /vgx?unofficialRun=43653414377。B200 数据点位于约 16.9 秒中位延迟处。

未验证

部署后的生产 /vgx、生产环境 v5 快照补建、带真实数据的 B300、zh 语言下的时长显示、InferenceX PR 评论中的 /vgx 链接（需在 InferenceX 仓库修改）。

AI disclosure: prepared with Perplexity Computer (orchestrator model Claude Opus 5.5).

_{Created with Perplexity}

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
inferencemax-app	Ready	Preview	Nov 15, 2026 1:01am UTC