feat(inference): add tokens per Calorie and Big Mac metrics / 新增每卡路里和每个 Big Mac 的 token 数指标 - #790
feat(inference): add tokens per Calorie and Big Mac metrics / 新增每卡路里和每个 Big Mac 的 token 数指标#790cquil11 wants to merge 5 commits into
Conversation
Add bilingual tokens-per-Calorie and tokens-per-Big-Mac Y-axis metrics, including historical trends, rooflines, unofficial-run overlays, and source-backed assumptions.\n\n中文：新增双语的每卡路里和每个 Big Mac 的 token 数 Y 轴指标，并接入历史趋势、Pareto 前沿、非官方运行覆盖及带来源的计算假设。
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Extend the Calorie and Big Mac axes with separate input-token and output-token variants across charts and historical trends.\n\n中文：为卡路里与 Big Mac 指标新增输入 token 和输出 token 变体，并接入推理图表与历史趋势。
Add labeled human output-speed guides to the output tokens per Calorie and output tokens per Big Mac charts. Keep the labels localized and aligned during chart zooms.\n\n中文：为每卡路里输出 token 数和每个 Big Mac 输出 token 数图表添加带标签的人类输出速度参考线，并确保标签支持中英文且在图表缩放时保持对齐。
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
There was a problem hiding this comment.
Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.
❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.
Reviewed by Cursor Bugbot for commit 34b9b9c. Configure here.
| const xScaleConfig = useStableValue(xScaleConfigRaw, isSameScaleConfig); | ||
|
|
||
| const yScaleConfigRaw = useMemo(() => { | ||
| const ext = |
There was a problem hiding this comment.
TTFT charts skip input scale logic
Medium Severity
y_inputTokensPerCalorie and y_inputTokensPerBigMac override the x-axis to P90 TTFT, and ChartControls already treats any “Input” title as an input metric (x-axis scale selector included). ScatterGraph still gates TTFT log/auto scaling and the y-log toggle on y_inputTputPerGpu only, so those new charts keep a linear x domain from 0 and ignore the scale control.
Additional Locations (2)
Reviewed by Cursor Bugbot for commit 34b9b9c. Configure here.
Keep the roofline and scatter layers at their original indices so tooltip hover continues to target scatter points. Render the human output guide later and lower it visually instead.\n\n中文：保持 roofline 和 scatter 图层的原始索引，使 tooltip 悬停继续正确绑定散点；人类输出参考线仍在后续渲染，并通过 SVG 层级置于数据后方。
Label both human output reference lines with the assumed average output rate from the discussion.\n\n中文：在两条人类输出参考线上标明讨论中采用的人类平均输出速度假设。
1 participant
Summary
Unofficial run support
Works for both official runs and
?unofficialrun=overlays through the sharedtransformBenchmarkRows→createChartDataPointpath. Preview verification was intentionally skipped for this quick meme PR.Validation
master(unrelated workspace export errors and an older oxlint ruleset)Sources
中文说明
非官方运行支持
官方运行与
?unofficialrun=非官方运行覆盖均通过共享的transformBenchmarkRows→createChartDataPoint数据路径获得这六个指标。本次为快速 meme PR，按要求未等待 Preview 并进行手动验证。验证
master，完整 typecheck 与 lint 未完成；当前报错为无关的 workspace 导出缺失及旧版 oxlint 规则集问题来源
Note
Low Risk
Additive display metrics and chart overlays on existing power/throughput data; no auth, persistence, or API contract changes.
Overview
Adds six food-energy Y-axis metrics (total / input / output tokens per food Calorie and per U.S. Big Mac) derived from all-in provisioned GPU power: 4,184 J per Calorie and 580 Cal per Big Mac.
Wires the fields through chart data points, Pareto rooflines, bilingual chart config, the Food Energy control group, and historical trend interpolation. Output-only calorie/Big Mac charts draw a dashed human output speed reference (3 tok/s) and expand the Y domain to include it. Captions cite NIST and McDonald’s nutrition sources.
Reviewed by Cursor Bugbot for commit 444b9aa. Bugbot is set up for automated code reviews on this repo. Configure here.