feat(power): add query energy and TDP dashboard axes / 新增每请求能耗与 TDP 占比图表轴

packages/app/cypress/component/inference-chart-controls.cy.tsx

@@ -41,6 +41,36 @@ describe('Inference ChartControls', () => {
     cy.get('@setSelectedYAxisMetric').should('have.been.calledOnce');
   });
 
+  it('lists and selects the schema-v2 derived axes in the Measured Energy group', () => {
+    const options = [
+      {
+        key: 'y_measuredJPerSuccessfulQuery',
+        label: 'Measured Joules per Successful Query',
+      },
+      {
+        key: 'y_measuredWhPerSuccessfulQuery',
+        label: 'Measured Watt-hours per Successful Query',
+      },
+      {
+        key: 'y_measuredPowerPercentTdp',
+        label: 'Measured Average Power as Percent of TDP',
+      },
+    ];
+
+    for (const option of options) {
+      cy.get('[data-testid="yaxis-metric-selector"]').click();
+      cy.contains('Measured Energy')
+        .parent()
+        .within(() => {
+          cy.contains('[role="option"]', option.label)
+            .scrollIntoView()
+            .should('be.visible')
+            .click();
+        });
+      cy.get('@setSelectedYAxisMetric').should('have.been.calledWith', option.key);
+    }
+  });
+
   it('hides the GPU comparison section when no GPUs are selected', () => {
     // Default mock: selectedGPUs = [] — GPU date range pickers should not render
     cy.contains('Comparison Date Range').should('not.exist');

packages/app/src/app/api/unofficial-run/route.test.ts

@@ -144,6 +144,55 @@ describe('normalizeArtifactRows', () => {
     expect(m.mean_e2el).toBe(1.5);
   });
 
+  it.each([
+    {
+      name: 'boolean verdict and junk-suffixed schema',
+      input: { power_valid: true, power_metric_schema_version: '2garbage' },
+      expectedVerdict: 0,
+      expectedSchema: undefined,
+    },
+    {
+      name: 'garbage verdict and valid numeric schema',
+      input: { power_valid: 'garbage', power_metric_schema_version: 2 },
+      expectedVerdict: 0,
+      expectedSchema: 2,
+    },
+    {
+      name: 'canonical numeric strings',
+      input: { power_valid: '1', power_metric_schema_version: '2' },
+      expectedVerdict: 1,
+      expectedSchema: 2,
+    },
+    {
+      name: 'explicit invalid verdict',
+      input: { power_valid: 0, power_metric_schema_version: 2 },
+      expectedVerdict: 0,
+      expectedSchema: 2,
+    },
+    {
+      name: 'legacy row without power contract fields',
+      input: {},
+      expectedVerdict: undefined,
+      expectedSchema: undefined,
+    },
+  ])(
+    'normalizes overlay power contract discriminators: $name',
+    ({ input, expectedVerdict, expectedSchema }) => {
+      const [row] = normalizeArtifactRows([rawRow(input)], '2026-03-01');
+
+      if (expectedVerdict === undefined) {
+        expect(row.metrics).not.toHaveProperty('power_valid');
+      } else {
+        expect(row.metrics.power_valid).toBe(expectedVerdict);
+      }
+      if (expectedSchema === undefined) {
+        expect(row.metrics).not.toHaveProperty('power_metric_schema_version');
+      } else {
+        expect(row.metrics.power_metric_schema_version).toBe(expectedSchema);
+      }
+    },
+  );
+
   it('surfaces pipeline-parallelism fields in metrics (auto-capture)', () => {
     // pp has no configs-table column: the frontend reads it from the metrics
     // JSONB (rowToAggDataEntry), so the overlay route must keep passing it

packages/app/src/components/ai-chart/types.test.ts

@@ -0,0 +1,15 @@
+import { describe, expect, it } from 'vitest';
+
+import { validateSpec } from './types';
+
+describe('validateSpec measured power axes', () => {
+  it.each([
+    'y_measuredJPerSuccessfulQuery',
+    'y_measuredWhPerSuccessfulQuery',
+    'y_measuredPowerPercentTdp',
+  ])('preserves %s as a benchmark Y-axis metric', (yAxisMetric) => {
+    const spec = validateSpec({ dataSource: 'benchmarks', yAxisMetric });
+
+    expect(spec.yAxisMetric).toBe(yAxisMetric);
+  });
+});

packages/app/src/components/inference/hooks/useInterpolatedTrendData.test.ts

@@ -156,6 +156,30 @@ describe('interpolateMetricAtInteractivity', () => {
     expect(result!).toBeLessThan(3.5);
   });
 
+  it.each([
+    'measuredJPerSuccessfulQuery',
+    'measuredWhPerSuccessfulQuery',
+    'measuredPowerPercentTdp',
+  ] as const)('interpolates the derived measured metric %s', (metricKey) => {
+    const points = [
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 800, roof: false },
+        [metricKey]: { y: 80, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 400, roof: false },
+        [metricKey]: { y: 40, roof: false },
+      }),
+    ];
+
+    const result = interpolateMetricAtInteractivity(points, 40, metricKey);
+    expect(result).not.toBeNull();
+    expect(result!).toBeGreaterThan(40);
+    expect(result!).toBeLessThan(80);
+  });
+
   it('returns null when metric field is missing from data points', () => {
     const points = [
       makePoint({ x: 20, tpPerGpu: { y: 800, roof: false } }),

packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts

@@ -11,7 +11,7 @@ import {
   reciprocalMetricAt,
 } from '@/components/calculator/useThroughputData';
 import { useBenchmarkHistory } from '@/hooks/api/use-benchmark-history';
-import { getHardwareKey } from '@/lib/chart-utils';
+import { buildMeasuredPowerChartFields, getHardwareKey } from '@/lib/chart-utils';
 import { getGpuSpecs, isKnownGpu } from '@/lib/constants';
 import { rowToAggDataEntry } from '@/lib/benchmark-transform';
 import type { BenchmarkRow } from '@/lib/api';
@@ -74,24 +74,7 @@ function rowToLightweightPoint(row: BenchmarkRow): InferenceData | null {
     jTotal: wrapMetric(power > 0 && tput ? (power * 1000) / tput : 0),
     ...(outputTput ? { jOutput: wrapMetric(power > 0 ? (power * 1000) / outputTput : 0) } : {}),
     ...(inputTput ? { jInput: wrapMetric(power > 0 ? (power * 1000) / inputTput : 0) } : {}),
-    ...(typeof entry.avg_power_w === 'number'
-      ? { measuredAvgPower: { y: entry.avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_output_token === 'number'
-      ? { measuredJPerOutputToken: { y: entry.joules_per_output_token, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_total_token === 'number'
-      ? { measuredJPerTotalToken: { y: entry.joules_per_total_token, roof: false } }
-      : {}),
-    ...(typeof entry.prefill_avg_power_w === 'number'
-      ? { measuredPrefillAvgPower: { y: entry.prefill_avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.decode_avg_power_w === 'number'
-      ? { measuredDecodeAvgPower: { y: entry.decode_avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_input_token === 'number'
-      ? { measuredJPerInputToken: { y: entry.joules_per_input_token, roof: false } }
-      : {}),
+    ...buildMeasuredPowerChartFields(entry, specs.tdp),
   };
   return point;
 }

packages/app/src/components/inference/inference-chart-config.json

@@ -164,6 +164,24 @@
     "y_measuredJPerTotalToken_title": "Measured Joules per Token (incl. prompt)",
     "y_measuredJPerTotalToken_titleZh": "每 token 实测焦耳能耗（含提示词）",
     "y_measuredJPerTotalToken_roofline": "lower_right",
+    "y_measuredJPerSuccessfulQuery": "measuredJPerSuccessfulQuery.y",
+    "y_measuredJPerSuccessfulQuery_label": "Measured J per Successful Query (J/query)",
+    "y_measuredJPerSuccessfulQuery_labelZh": "每次成功请求实测能耗（J/query）",
+    "y_measuredJPerSuccessfulQuery_title": "Measured Joules per Successful Query",
+    "y_measuredJPerSuccessfulQuery_titleZh": "每次成功请求实测焦耳能耗",
+    "y_measuredJPerSuccessfulQuery_roofline": "lower_right",
+    "y_measuredWhPerSuccessfulQuery": "measuredWhPerSuccessfulQuery.y",
+    "y_measuredWhPerSuccessfulQuery_label": "Measured Wh per Successful Query (Wh/query)",
+    "y_measuredWhPerSuccessfulQuery_labelZh": "每次成功请求实测能耗（Wh/query）",
+    "y_measuredWhPerSuccessfulQuery_title": "Measured Watt-hours per Successful Query",
+    "y_measuredWhPerSuccessfulQuery_titleZh": "每次成功请求实测瓦时能耗",
+    "y_measuredWhPerSuccessfulQuery_roofline": "lower_right",
+    "y_measuredPowerPercentTdp": "measuredPowerPercentTdp.y",
+    "y_measuredPowerPercentTdp_label": "Measured Average Power (% TDP)",
+    "y_measuredPowerPercentTdp_labelZh": "实测平均功耗（TDP 占比）",
+    "y_measuredPowerPercentTdp_title": "Measured Average Power as Percent of TDP",
+    "y_measuredPowerPercentTdp_titleZh": "实测平均功耗占 TDP 百分比",
+    "y_measuredPowerPercentTdp_roofline": "lower_right",
     "y_cost_limit": 5,
     "y_latency_limit": 60
   },
@@ -331,6 +349,24 @@
     "y_measuredJPerTotalToken_title": "Measured Joules per Token (incl. prompt)",
     "y_measuredJPerTotalToken_titleZh": "每 token 实测焦耳能耗（含提示词）",
     "y_measuredJPerTotalToken_roofline": "lower_left",
+    "y_measuredJPerSuccessfulQuery": "measuredJPerSuccessfulQuery.y",
+    "y_measuredJPerSuccessfulQuery_label": "Measured J per Successful Query (J/query)",
+    "y_measuredJPerSuccessfulQuery_labelZh": "每次成功请求实测能耗（J/query）",
+    "y_measuredJPerSuccessfulQuery_title": "Measured Joules per Successful Query",
+    "y_measuredJPerSuccessfulQuery_titleZh": "每次成功请求实测焦耳能耗",
+    "y_measuredJPerSuccessfulQuery_roofline": "lower_left",
+    "y_measuredWhPerSuccessfulQuery": "measuredWhPerSuccessfulQuery.y",
+    "y_measuredWhPerSuccessfulQuery_label": "Measured Wh per Successful Query (Wh/query)",
+    "y_measuredWhPerSuccessfulQuery_labelZh": "每次成功请求实测能耗（Wh/query）",
+    "y_measuredWhPerSuccessfulQuery_title": "Measured Watt-hours per Successful Query",
+    "y_measuredWhPerSuccessfulQuery_titleZh": "每次成功请求实测瓦时能耗",
+    "y_measuredWhPerSuccessfulQuery_roofline": "lower_left",
+    "y_measuredPowerPercentTdp": "measuredPowerPercentTdp.y",
+    "y_measuredPowerPercentTdp_label": "Measured Average Power (% TDP)",
+    "y_measuredPowerPercentTdp_labelZh": "实测平均功耗（TDP 占比）",
+    "y_measuredPowerPercentTdp_title": "Measured Average Power as Percent of TDP",
+    "y_measuredPowerPercentTdp_titleZh": "实测平均功耗占 TDP 百分比",
+    "y_measuredPowerPercentTdp_roofline": "lower_left",
     "y_cost_limit": 5,
     "y_latency_limit": 60
   }

packages/app/src/components/inference/measured-power-direction.test.ts

@@ -30,6 +30,12 @@ const MEASURED_POWER_METRICS = [
   'y_measuredDecodeAvgPower',
 ] as const;
 
+const DERIVED_POWER_METRICS = [
+  'y_measuredJPerSuccessfulQuery',
+  'y_measuredWhPerSuccessfulQuery',
+  'y_measuredPowerPercentTdp',
+] as const;
+
 const defs = chartDefinitions as unknown as ChartDefinition[];
 const interactivityDef = defs.find((d) => d.chartType === 'interactivity')!;
 const e2eDef = defs.find((d) => d.chartType === 'e2e')!;
@@ -133,4 +139,13 @@ describe('measured-power Pareto direction', () => {
       }
     }
   });
+
+  it.each(DERIVED_POWER_METRICS)('%s is bilingual and lower-is-better', (metric) => {
+    for (const chartDef of [interactivityDef, e2eDef]) {
+      expect(chartDef[metric]).toMatch(/\.y$/u);
+      expect(chartDef[`${metric}_label`]).toBeTruthy();
+      expect(chartDef[`${metric}_labelZh`]).toBeTruthy();
+      expect(declaredDirection(chartDef, metric)).toMatch(/^lower_/u);
+    }
+  });
 });

packages/app/src/components/inference/types.ts

@@ -137,21 +137,24 @@ export interface AggDataEntry {
   'p99.9_e2el': number;
   // Measured GPU telemetry (emitted by runner's aggregate_power.py).
   // Optional because historical runs predate the fields.
+  power_valid?: number;
+  power_metric_schema_version?: number;
   avg_power_w?: number;
+  joules_per_successful_query?: number;
   joules_per_output_token?: number;
   joules_per_total_token?: number;
   // Multinode / disagg-only measured power. The aggregate_power.py runner
   // emits per-role energy splits when the deployment has separate prefill
   // and decode workers (single-node disagg or multinode disagg). Single-node
   // aggregated configs leave these undefined.
   // - prefill_avg_power_w / decode_avg_power_w: mean per-GPU draw (W) within each role
-  // - joules_per_input_token: prefill_energy / total_input_tokens (prefill GPUs only)
-  // The disagg decode-only J/output is carried by joules_per_output_token above
-  // (the runner overrides it to decode_energy / total_output_tokens on disagg) —
-  // there is no separate _decode field.
+  // Unprefixed joules fields are whole-deployment metrics in schema version 2.
+  // Explicit prefill/decode keys retain role-local energy breakdowns.
   prefill_avg_power_w?: number;
   decode_avg_power_w?: number;
   joules_per_input_token?: number;
+  prefill_joules_per_input_token?: number;
+  decode_joules_per_output_token?: number;
   // Cluster-wide GPU telemetry beyond power (temperature, utilization, memory).
   // Emitted by aggregate_power.py when the perfmon CSVs include the matching
   // sample columns. Optional because older runs (and runs without the relevant
@@ -315,6 +318,9 @@ export interface InferenceData extends Partial<Omit<AggDataEntry, AggDataConflic
   measuredJPerOutputToken?: { y: number; roof: boolean };
   measuredJPerTotalToken?: { y: number; roof: boolean };
   measuredJPerInputToken?: { y: number; roof: boolean };
+  measuredJPerSuccessfulQuery?: { y: number; roof: boolean };
+  measuredWhPerSuccessfulQuery?: { y: number; roof: boolean };
+  measuredPowerPercentTdp?: { y: number; roof: boolean };
 }
 
 /** Why a chart-ready point was intentionally excluded from the visible plot. */
@@ -358,7 +364,10 @@ export type YAxisMetricKey =
   | 'measuredDecodeAvgPower'
   | 'measuredJPerOutputToken'
   | 'measuredJPerTotalToken'
-  | 'measuredJPerInputToken';
+  | 'measuredJPerInputToken'
+  | 'measuredJPerSuccessfulQuery'
+  | 'measuredWhPerSuccessfulQuery'
+  | 'measuredPowerPercentTdp';
 
 /**
  * Defines the configuration and labels for a specific chart.
@@ -488,6 +497,26 @@ export interface ChartDefinition {
   y_measuredJPerTotalToken_label?: string;
   y_measuredJPerTotalToken_title?: string;
   y_measuredJPerTotalToken_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
+  y_measuredJPerSuccessfulQuery?: string;
+  y_measuredJPerSuccessfulQuery_label?: string;
+  y_measuredJPerSuccessfulQuery_title?: string;
+  y_measuredJPerSuccessfulQuery_roofline?:
+    | 'upper_right'
+    | 'upper_left'
+    | 'lower_left'
+    | 'lower_right';
+  y_measuredWhPerSuccessfulQuery?: string;
+  y_measuredWhPerSuccessfulQuery_label?: string;
+  y_measuredWhPerSuccessfulQuery_title?: string;
+  y_measuredWhPerSuccessfulQuery_roofline?:
+    | 'upper_right'
+    | 'upper_left'
+    | 'lower_left'
+    | 'lower_right';
+  y_measuredPowerPercentTdp?: string;
+  y_measuredPowerPercentTdp_label?: string;
+  y_measuredPowerPercentTdp_title?: string;
+  y_measuredPowerPercentTdp_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
   y_cost_limit?: number;
   y_latency_limit?: number;
 }

packages/app/src/components/inference/ui/ChartControls.tsx

@@ -157,6 +157,9 @@ const METRIC_GROUPS: {
       'y_measuredJPerInputToken',
       'y_measuredJPerOutputToken',
       'y_measuredJPerTotalToken',
+      'y_measuredJPerSuccessfulQuery',
+      'y_measuredWhPerSuccessfulQuery',
+      'y_measuredPowerPercentTdp',
     ],
   },
   { label: 'Custom User Values', labelZh: '自定义值', metrics: ['y_costUser', 'y_powerUser'] },