feat(dashboard): surface prefill/decode power-per-GPU + J/Input (measured energy)

[arygupt]

What

Adds Prefill Power/GPU, Decode Power/GPU, and J / Input token as selectable metrics in the Measured-Energy group (behind the existing konami gate), alongside the avg-power / J-out / J-total metrics already on master.

Re-land of #411 (closed unmerged) — cherry-picked cleanly onto a fresh branch off master:

• d63103c feat(dashboard): add prefill/decode/J-input as measured-energy metrics
• 3f2399c style(chart-utils): satisfy oxfmt format check

Why

The per-stage measured power these metrics read is now live in prod. The multinode disagg rows — gb300 DeepSeek-V4-Pro and mi355x MiniMax-M2.5 — carry prefill_avg_power_w / decode_avg_power_w / joules_per_input_token as of today's ingest (run 26607091549). This PR is the display layer that surfaces them; without it the data is queryable but not selectable in the chart/table.

Scope / safety

• Pure frontend: adds chart-definition entries + labels only. No query, schema, or ingest changes — those landed in feat(power): measured-power multinode support (workers[] + per-stage joules) #405.
• Vendor-agnostic: gb300 and AMD ride the same scatter/table path.
• Gated behind the Measured-Energy toggle, so no change to the default public view.

Test plan

• CI green (lint / typecheck / unit / build)
• Metrics appear in the Measured-Energy dropdown; selecting Prefill/Decode Power renders the table column
• Local: tsc clean, unit tests green

🤖 Generated with Claude Code

---

Note

Low Risk
Pure frontend chart/config and type wiring with optional fields and existing konami gate; no API, ingest, or auth changes.

Overview
Surfaces per-stage measured energy in the inference dashboard: prefill/decode average power per GPU and measured J per input token, alongside the existing cluster-average power and J/out/J/total metrics.

Chart points now map runner fields prefill_avg_power_w, decode_avg_power_w, and joules_per_input_token into optional measuredPrefillAvgPower, measuredDecodeAvgPower, and measuredJPerInputToken (same typeof === 'number' gating as other telemetry so legacy rows stay absent, not zero). Wiring is mirrored in createChartDataPoint, trend interpolation (useInterpolatedTrendData), inference-chart-config.json (interactivity + e2e), roofline unions, and the Measured Energy Y-axis group in ChartControls (still feature-gated).

Unit tests cover independent prefill/decode emission, legacy omission, and zero-W preservation.

^{Reviewed by Cursor Bugbot for commit 3f2399c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
inferencemax-app	Ready	Preview, Comment	Jun 1, 2026 8:24pm

[claude]

Claude finished @arygupt's task in 2m 18s —— View job

---

Review complete

🟢 LGTM — no blocking issues found

What I checked

• chart-utils.ts — all 4 metric registration sites updated (Y_AXIS_METRICS, createChartDataPoint, calculateRoofline/computeAllRooflines type unions, markRooflinePoints reset + assign)
• types.ts — InferenceData, YAxisMetricKey, and ChartDefinition all extended symmetrically
• inference-chart-config.json — entries added to both e2e and interactivity chart definitions, with roofline directions for JPerInputToken matching the JPerOutputToken pattern (lower_right / lower_left)
• ChartControls.tsx — new metrics added to the Measured-Energy group (behind the existing gated: true konami toggle)
• useInterpolatedTrendData.ts — rowToLightweightPoint mirrors the new createChartDataPoint branches
• Tests — 7 new unit tests cover each field independently, the legacy-row omission case, zero-value preservation (vs. falsy coercion), and a combined full-row check
• Overlay path covered: buildChartData → transformBenchmarkRows → rowToAggDataEntry (already carries prefill_avg_power_w / decode_avg_power_w / joules_per_input_token) → createChartDataPoint. No overlay-specific code branch needed for this kind of metric-registration-only change.

Notes

• The two new per-GPU power metrics intentionally have no _roofline config key, matching measuredAvgPower — power-per-GPU has no meaningful frontier. computeAllRooflines guards on actualDataYKey && rooflineDirection, so this skips cleanly.
• benchmark-transform.test.ts already has coverage for the BenchmarkRow → AggDataEntry side of these three fields; combined with the new chart-utils tests, both transform stages are tested.
• No new interactive elements introduced — the existing metric-dropdown change tracking already fires when these options are selected, so no new track() call is needed.