feat(dashboard): render per-prefill / per-decode / total measured power

[arygupt]

What

Adds three new selectable Y-axis metrics to the inference chart, surfacing the per-stage measured-power telemetry the runner already emits but the dashboard could not render:

New metric	Source field (already on AggDataEntry)	Mirrors
Measured Prefill Power per GPU (W)	prefill_avg_power_w	measuredAvgPower
Measured Decode Power per GPU (W)	decode_avg_power_w	measuredAvgPower
Measured J per Input Token (J/tok)	joules_per_input_token	measuredJPerOutputToken

Net result in the gated "Measured Energy" dropdown: prefill / decode / total power + J input / output / total — i.e. the per-prefill/per-decode/per-total split for disaggregated runs.

Why

The disagg per-stage power data flows all the way through the pipeline (DB metrics JSONB + unofficial-run artifact path → rowToAggDataEntry → chart point) but only three measured metrics were ever wired as selectable Y-axis options (total power, J/output, J/total). The per-stage fields existed on the row but were never wrapped into metrics. This closes that gap.

How

Purely additive / mechanical — copies the existing measuredAvgPower trio pattern across every site:

• inference-chart-config.json (both interactivity + e2e charts; J/input roofline lower_right/lower_left to match J/output)
• Y_AXIS_METRICS, YAxisMetricKey, ChartDefinition, InferenceData (types)
• createChartDataPoint + roofline machinery (both type unions, roof-reset, markRooflinePoints)
• useInterpolatedTrendData lightweight trend-point builder
• ChartControls gated "Measured Energy" group (stays gated: true)
• chart-utils.test.ts — 7 new cases (emit / omit-legacy / zero-preservation / per-stage independence / full-disagg row)

No runner, ETL, or DB change — source fields already exist on AggDataEntry; packages/constants already lists the keys.

Still gated

The metrics remain behind the existing ↑ ↑ ↓ ↓ feature gate, so merging does not expose measured power publicly. Ungating is a separate, deliberate product decision (follow-up).

How to preview (no DB needed)

On the Vercel preview this PR posts:

1. Open …/inference?unofficialrun=26607091549 (case-insensitive; fetches the GB300 run's GitHub artifacts directly — needs GITHUB_TOKEN in the preview env, which already powers the PR unofficial-run visualizer).
2. Press ↑ ↑ ↓ ↓ to unlock "Measured Energy".
3. In the Y-axis dropdown pick Measured Prefill Power per GPU, then Measured Decode Power per GPU, then Measured Average Power per GPU; filter to the dsv4/gb300 points → screenshot each. Expect prefill > decode per-GPU watts (compute- vs memory-bound).

Verification

• tsc --noEmit — clean (the Y_AXIS_METRICS/YAxisMetricKey/ChartDefinition/InferenceData sites all agree).
• Full app unit suite — 1996 passed (incl. 7 new cases).
• oxlint on changed files — clean.

Out of scope (follow-ups)

• Combined prefill+decode+total overlay as one multi-series chart.
• Per-worker workers[] visualization (data is carried to the point, not yet rendered).
• Ungating measured power for the public site.

🤖 Generated with Claude Code

---

Note

Low Risk
Additive frontend chart and type wiring with optional-field guards; no API, auth, or data-pipeline changes.

Overview
Adds three selectable measured-energy Y-axis metrics on the inference dashboard—prefill power, decode power, and J per input token—so disaggregated runs can show per-stage telemetry that already exists on benchmark rows but was not chartable.

The change mirrors the existing total measured-power pattern: map prefill_avg_power_w, decode_avg_power_w, and joules_per_input_token into chart points in createChartDataPoint and trend interpolation, register them in chart config and Y_AXIS_METRICS, extend roofline marking, and list them under the gated Measured Energy dropdown (prefill/decode/total power plus input/output/total J). Legacy rows without per-stage fields stay omitted (not zero-filled). Unit tests cover emit, omit, zero preservation, and full disagg rows.

No runner, DB, or ETL changes; the feature gate for measured energy is unchanged.

^{Reviewed by Cursor Bugbot for commit 0e85ae7. 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 3:09am

[claude]

Claude finished @arygupt's task in 2m 40s —— View job

---

🟢 LGTM — no blocking issues found

This is a clean, mechanical extension of the existing measured-power trio. I verified every sync point ends up consistent:

Sync points checked (all 11 in agreement):

• Y_AXIS_METRICS + YAxisMetric type (chart-utils.ts:151-160)
• YAxisMetricKey (types.ts:265-270)
• ChartDefinition field set (types.ts:379-394)
• InferenceData wrapper fields (types.ts:233-238)
• createChartDataPoint source-field guards (chart-utils.ts:407-422)
• calculateRoofline / computeAllRooflines union types (chart-utils.ts:581-663)
• markRooflinePoints reset block + if/else chain (chart-utils.ts:706-795)
• useInterpolatedTrendData lightweight builder (useInterpolatedTrendData.ts:83-91)
• inference-chart-config.json (both interactivity + e2e blocks; J/input roofline = lower_right / lower_left to mirror J/output)
• ChartControls.tsx Measured Energy group (gated: true preserved)
• Tests: 7 new cases in chart-utils.test.ts:1297-1369 cover emit / omit-legacy / zero-preservation / per-stage independence / full-disagg row

Unofficial-run path: Source fields (prefill_avg_power_w, decode_avg_power_w, joules_per_input_token) are already mapped by benchmark-transform.ts:61-63 (covered by benchmark-transform.test.ts), and the unofficial-run provider uses the same transformBenchmarkRows. So overlay rooflines/points pick up the new metrics for free — overlay support is structurally present.

typeof === 'number' guards preserve 0 W values (explicitly asserted by the new "preserves a zero per-stage power value" test).

Feature gate stays in place — merging does not expose measured power publicly.

Minor pre-existing observation (not introduced by this PR, not blocking): chart-display-helpers.tsx:117 uses startsWith('y_j') to gate the Joules disagg caveat, which doesn't match any y_measured* metric — including the existing y_measuredJPerOutputToken/y_measuredJPerTotalToken. Since the new y_measuredJPerInputToken follows the same pre-existing pattern, there's no new regression here. If/when the Measured Energy group is ungated, that's the place where a y_measured caveat branch would need to be added.

• Read changed files
• Verify mechanical-copy claim across 11 sync points
• Check unofficial-run path coverage
• Check tests (7 new cases)
• Post review
  • Branch: feat/dashboard-prefill-decode-power