@simulatte/doppler 0.1.3 → 0.1.5

package/src/inference/browser-harness.js

@@ -15,10 +15,16 @@ import {
   getActiveKernelPathSource,
   getActiveKernelPathPolicy,
 } from '../config/kernel-path-loader.js';
-import { selectRuleValue } from '../rules/rule-registry.js';
+import {
+  getInferenceLayerPatternContractArtifact,
+  selectRuleValue,
+} from '../rules/rule-registry.js';
 import { mergeRuntimeValues } from '../config/runtime-merge.js';
 import { isPlainObject } from '../utils/plain-object.js';
+import { validateBrowserSuiteMetrics } from '../config/schema/browser-suite-metrics.schema.js';
 import { validateTrainingMetricsReport } from '../config/schema/training-metrics.schema.js';
+import { buildExecutionContractArtifact } from '../config/execution-contract-check.js';
+import { buildManifestRequiredInferenceFieldsArtifact } from '../config/required-inference-fields-contract-check.js';
 
 const TRAINING_SUITE_MODULE_PATH = '../training/suite.js';
 const NODE_SOURCE_RUNTIME_MODULE_PATH = '../tooling/node-source-runtime.js';
@@ -41,6 +47,29 @@ async function runTrainingBenchSuite(options = {}) {
   return module.runTrainingBenchSuite(options);
 }
 
+function buildSuiteContractMetrics(suite, baseMetrics, manifest) {
+  const executionContractArtifact = buildExecutionContractArtifact(manifest);
+  const executionV0GraphContractArtifact = executionContractArtifact?.executionV0?.graph ?? null;
+  const layerPatternContractArtifact = getInferenceLayerPatternContractArtifact();
+  const requiredInferenceFieldsArtifact = manifest?.modelType === 'transformer'
+    && isPlainObject(manifest?.inference?.attention)
+    ? buildManifestRequiredInferenceFieldsArtifact(
+      manifest?.inference ?? null,
+      `${manifest?.modelId ?? 'unknown'}.inference`
+    )
+    : null;
+  return validateBrowserSuiteMetrics({
+    ...baseMetrics,
+    schemaVersion: 1,
+    source: 'doppler',
+    suite,
+    ...(executionContractArtifact ? { executionContractArtifact } : {}),
+    executionV0GraphContractArtifact,
+    layerPatternContractArtifact,
+    requiredInferenceFieldsArtifact,
+  });
+}
+
 function parseReportTimestamp(rawTimestamp, label = 'timestamp') {
   if (rawTimestamp == null) {
     return null;
@@ -958,6 +987,22 @@ async function runKernelSuite(options = {}) {
 
 
 const DEFAULT_HARNESS_PROMPT = 'Summarize this input in one sentence.';
+const DEFAULT_RUNTIME_PLACEHOLDER_PROMPT = 'Hello from Doppler.';
+const DEFAULT_TRANSLATEGEMMA_PROMPT = Object.freeze({
+  messages: Object.freeze([
+    Object.freeze({
+      role: 'user',
+      content: Object.freeze([
+        Object.freeze({
+          type: 'text',
+          source_lang_code: 'en',
+          target_lang_code: 'fr',
+          text: 'Hello world.',
+        }),
+      ]),
+    }),
+  ]),
+});
 const DEFAULT_HARNESS_MAX_TOKENS = 32;
 const EMBEDDING_PREVIEW_LENGTH = 16;
 const EMBEDDING_SEMANTIC_MIN_RETRIEVAL_TOP1 = 0.67;
@@ -1205,24 +1250,90 @@ function isStructuredPromptInput(value) {
   return Array.isArray(value) || (value != null && typeof value === 'object');
 }
 
-function resolveGenerationPromptInput(runtimeConfig, runOverrides = null) {
+function clonePromptInput(promptInput) {
+  if (!isStructuredPromptInput(promptInput)) {
+    return promptInput;
+  }
+  if (typeof structuredClone === 'function') {
+    return structuredClone(promptInput);
+  }
+  return JSON.parse(JSON.stringify(promptInput));
+}
+
+function resolvePromptTemplateType(source) {
+  const sourceTemplateType = asText(source?.chatTemplateType);
+  if (sourceTemplateType) {
+    return sourceTemplateType;
+  }
+  const modelConfigTemplateType = asText(source?.modelConfig?.chatTemplateType);
+  if (modelConfigTemplateType) {
+    return modelConfigTemplateType;
+  }
+  return asText(source?.manifest?.inference?.chatTemplate?.type);
+}
+
+function buildDefaultGenerationPrompt(templateType) {
+  if (templateType === 'translategemma') {
+    return clonePromptInput(DEFAULT_TRANSLATEGEMMA_PROMPT);
+  }
+  return DEFAULT_HARNESS_PROMPT;
+}
+
+function shouldPreferModelDefaultPrompt(runtimePrompt, templateType) {
+  if (templateType !== 'translategemma') {
+    return false;
+  }
+  if (typeof runtimePrompt !== 'string') {
+    return false;
+  }
+  return runtimePrompt.trim() === DEFAULT_RUNTIME_PLACEHOLDER_PROMPT;
+}
+
+function describePromptInput(promptInput) {
+  if (typeof promptInput === 'string') {
+    return promptInput.trim() || DEFAULT_HARNESS_PROMPT;
+  }
+  const firstMessage = Array.isArray(promptInput?.messages)
+    ? promptInput.messages[0]
+    : null;
+  const firstContent = Array.isArray(firstMessage?.content)
+    ? firstMessage.content[0]
+    : null;
+  const sourceLang = asText(firstContent?.source_lang_code);
+  const targetLang = asText(firstContent?.target_lang_code);
+  const text = asText(firstContent?.text);
+  if (sourceLang && targetLang) {
+    return `${sourceLang} -> ${targetLang}: ${text || '[non-text request]'}`;
+  }
+  try {
+    return JSON.stringify(promptInput);
+  } catch {
+    return '[structured prompt]';
+  }
+}
+
+function resolveGenerationPromptInput(runtimeConfig, runOverrides = null, source = null) {
+  const templateType = resolvePromptTemplateType(source);
   const overridePrompt = runOverrides?.prompt;
   if (typeof overridePrompt === 'string' && overridePrompt.trim()) {
     return overridePrompt.trim();
   }
   if (isStructuredPromptInput(overridePrompt)) {
-    return overridePrompt;
+    return clonePromptInput(overridePrompt);
   }
 
   const runtimePrompt = runtimeConfig?.inference?.prompt;
+  if (shouldPreferModelDefaultPrompt(runtimePrompt, templateType)) {
+    return buildDefaultGenerationPrompt(templateType);
+  }
   if (typeof runtimePrompt === 'string' && runtimePrompt.trim()) {
     return runtimePrompt.trim();
   }
   if (isStructuredPromptInput(runtimePrompt)) {
-    return runtimePrompt;
+    return clonePromptInput(runtimePrompt);
   }
 
-  return DEFAULT_HARNESS_PROMPT;
+  return buildDefaultGenerationPrompt(templateType);
 }
 
 function resolveMaxTokens(runtimeConfig) {
@@ -1233,7 +1344,7 @@ function resolveMaxTokens(runtimeConfig) {
   return DEFAULT_HARNESS_MAX_TOKENS;
 }
 
-function resolveBenchmarkRunSettings(runtimeConfig) {
+function resolveBenchmarkRunSettings(runtimeConfig, source = null) {
   const benchConfig = runtimeConfig?.shared?.benchmark?.run || {};
   const runtimeSampling = isPlainObject(runtimeConfig?.inference?.sampling)
     ? runtimeConfig.inference.sampling
@@ -1241,9 +1352,9 @@ function resolveBenchmarkRunSettings(runtimeConfig) {
   const benchSampling = isPlainObject(benchConfig?.sampling)
     ? benchConfig.sampling
     : {};
-  const prompt = typeof benchConfig.customPrompt === 'string' && benchConfig.customPrompt.trim()
+  const promptInput = typeof benchConfig.customPrompt === 'string' && benchConfig.customPrompt.trim()
     ? benchConfig.customPrompt.trim()
-    : resolvePrompt(runtimeConfig);
+    : resolveGenerationPromptInput(runtimeConfig, null, source);
   const maxTokens = Number.isFinite(benchConfig.maxNewTokens)
     ? Math.max(1, Math.floor(benchConfig.maxNewTokens))
     : resolveMaxTokens(runtimeConfig);
@@ -1251,7 +1362,8 @@ function resolveBenchmarkRunSettings(runtimeConfig) {
   return {
     warmupRuns: Math.max(0, Math.floor(benchConfig.warmupRuns ?? 0)),
     timedRuns: Math.max(1, Math.floor(benchConfig.timedRuns ?? 1)),
-    prompt,
+    prompt: promptInput,
+    promptLabel: describePromptInput(promptInput),
     maxTokens,
     sampling: {
       ...runtimeSampling,
@@ -1465,7 +1577,8 @@ function isCoherentOutput(tokens, output) {
 async function runGeneration(pipeline, runtimeConfig, runOverrides = null) {
   const tokens = [];
   const tokenIds = [];
-  const promptInput = resolveGenerationPromptInput(runtimeConfig, runOverrides);
+  const promptInput = resolveGenerationPromptInput(runtimeConfig, runOverrides, pipeline);
+  const promptLabel = describePromptInput(promptInput);
   const useChatTemplate = runOverrides?.useChatTemplate
     ?? runtimeConfig?.inference?.chatTemplate?.enabled
     ?? (isStructuredPromptInput(promptInput) ? true : undefined);
@@ -1532,7 +1645,8 @@ async function runGeneration(pipeline, runtimeConfig, runOverrides = null) {
     : null;
 
   return {
-    prompt: promptInput,
+    prompt: promptLabel,
+    promptInput,
     maxTokens,
     tokens,
     tokenIds,
@@ -1739,6 +1853,11 @@ async function runInferenceSuite(options = {}) {
     source: 'doppler',
     prefillSemantics: 'internal_prefill_phase',
   });
+  const metricsWithContracts = buildSuiteContractMetrics(
+    options.suiteName || 'inference',
+    metrics,
+    harness.manifest
+  );
   return {
     ...summary,
     modelId: options.modelId || harness.manifest?.modelId || 'unknown',
@@ -1756,7 +1875,7 @@ async function runInferenceSuite(options = {}) {
     timing,
     timingDiagnostics,
     output,
-    metrics,
+    metrics: metricsWithContracts,
     memoryStats,
     deviceInfo: resolveDeviceInfo(),
     pipeline: options.keepPipeline ? harness.pipeline : null,
@@ -1766,9 +1885,9 @@ async function runInferenceSuite(options = {}) {
 async function runBenchSuite(options = {}) {
   const startTime = performance.now();
   const runtimeConfig = getRuntimeConfig();
-  const benchRun = resolveBenchmarkRunSettings(runtimeConfig);
-  const warmupRuns = benchRun.warmupRuns;
-  const timedRuns = benchRun.timedRuns;
+  const defaultBenchRun = resolveBenchmarkRunSettings(runtimeConfig);
+  const warmupRuns = defaultBenchRun.warmupRuns;
+  const timedRuns = defaultBenchRun.timedRuns;
   const cacheMode = normalizeCacheMode(options.cacheMode);
   const loadMode = normalizeLoadMode(options.loadMode, !options.modelUrl);
   const workloadType = normalizeWorkloadType(options.workloadType);
@@ -1776,7 +1895,7 @@ async function runBenchSuite(options = {}) {
   if (workloadType === 'training') {
     const trainingBench = await runTrainingBenchSuite({
       ...options,
-      benchRun,
+      benchRun: defaultBenchRun,
       workloadType,
     });
     const trainingReport = trainingBench?.metrics?.trainingMetricsReport;
@@ -1858,6 +1977,7 @@ async function runBenchSuite(options = {}) {
   }
 
   const harness = await initializeSuiteModel(options);
+  const benchRun = resolveBenchmarkRunSettings(runtimeConfig, harness.pipeline ?? harness);
   const modelType = harness.manifest?.modelType || 'transformer';
   const safeModelLoadMs = toTimingNumber(harness.modelLoadMs, 0);
 
@@ -1927,7 +2047,7 @@ async function runBenchSuite(options = {}) {
       validRuns: durations.length,
       invalidRuns,
       invalidRatePct: Number((timedRuns > 0 ? (invalidRuns / timedRuns) * 100 : 0).toFixed(2)),
-      prompt: benchRun.prompt,
+      prompt: benchRun.promptLabel,
       embeddingDim: Math.round(embeddingDims.reduce((a, b) => a + b, 0) / (embeddingDims.length || 1)),
       nonFiniteValues: totalNonFiniteValues,
       firstTimedEmbeddingMs: Number((firstTimedEmbeddingMs ?? 0).toFixed(2)),
@@ -2058,7 +2178,7 @@ async function runBenchSuite(options = {}) {
     metrics = {
       warmupRuns,
       timedRuns,
-      prompt: benchRun.prompt,
+      prompt: benchRun.promptLabel,
       maxTokens: benchRun.maxTokens,
       decodeTokensPerSec: sampleTimingNumber(decodeTokensPerSecStats, 'median'),
       avgTokensGenerated: Math.round(tokensGeneratedStats.mean),
@@ -2132,6 +2252,7 @@ async function runBenchSuite(options = {}) {
     source: 'doppler',
     prefillSemantics: 'internal_prefill_phase',
   });
+  const metricsWithContracts = buildSuiteContractMetrics('bench', metrics, harness.manifest);
   return {
     ...summary,
     modelId: options.modelId || harness.manifest?.modelId || 'unknown',
@@ -2149,7 +2270,7 @@ async function runBenchSuite(options = {}) {
     timing,
     timingDiagnostics,
     output,
-    metrics,
+    metrics: metricsWithContracts,
     memoryStats,
     deviceInfo: resolveDeviceInfo(),
     pipeline: options.keepPipeline ? harness.pipeline : null,
@@ -2310,25 +2431,9 @@ async function runDiffusionSuite(options = {}) {
     source: 'doppler',
     prefillSemantics: 'internal_prefill_phase',
   });
-
-  return {
-    ...summary,
-    modelId: options.modelId || harness.manifest?.modelId || 'unknown',
-    cacheMode,
-    loadMode,
-    env: {
-      library: 'doppler',
-      runtime: 'browser',
-      device: 'webgpu',
-      browserUserAgent: typeof navigator !== 'undefined' ? (navigator.userAgent || null) : null,
-      browserPlatform: typeof navigator !== 'undefined' ? (navigator.platform || null) : null,
-      browserLanguage: typeof navigator !== 'undefined' ? (navigator.language || null) : null,
-      browserVendor: typeof navigator !== 'undefined' ? (navigator.vendor || null) : null,
-    },
-    timing,
-    timingDiagnostics,
-    output,
-    metrics: {
+  const metricsWithContracts = buildSuiteContractMetrics(
+    'diffusion',
+    {
       warmupRuns,
       timedRuns,
       width,
@@ -2353,6 +2458,27 @@ async function runDiffusionSuite(options = {}) {
       gpu: gpuStats,
       performanceArtifact: diffusionPerformanceArtifact,
     },
+    harness.manifest
+  );
+
+  return {
+    ...summary,
+    modelId: options.modelId || harness.manifest?.modelId || 'unknown',
+    cacheMode,
+    loadMode,
+    env: {
+      library: 'doppler',
+      runtime: 'browser',
+      device: 'webgpu',
+      browserUserAgent: typeof navigator !== 'undefined' ? (navigator.userAgent || null) : null,
+      browserPlatform: typeof navigator !== 'undefined' ? (navigator.platform || null) : null,
+      browserLanguage: typeof navigator !== 'undefined' ? (navigator.language || null) : null,
+      browserVendor: typeof navigator !== 'undefined' ? (navigator.vendor || null) : null,
+    },
+    timing,
+    timingDiagnostics,
+    output,
+    metrics: metricsWithContracts,
     memoryStats,
     deviceInfo: resolveDeviceInfo(),
     pipeline: options.keepPipeline ? harness.pipeline : null,

package/src/inference/pipelines/diffusion/init.js

@@ -1,5 +1,7 @@
 import { DEFAULT_DIFFUSION_CONFIG } from '../../../config/schema/index.js';
 
+const SUPPORTED_DIFFUSION_RUNTIME_LAYOUTS = new Set(['sd3', 'flux', 'sana']);
+
 function mergeSection(base, override) {
   if (!override) return { ...base };
   return { ...base, ...override };
@@ -38,6 +40,9 @@ function resolveSchedulerType(modelScheduler, runtimeScheduler) {
   if (modelClass === 'FlowMatchEulerDiscreteScheduler') {
     return 'flowmatch_euler';
   }
+  if (modelClass === 'SCMScheduler') {
+    return 'scm';
+  }
   if (modelClass === 'EulerDiscreteScheduler') {
     return 'euler';
   }
@@ -58,6 +63,8 @@ function mergeSchedulerConfig(modelConfig, runtimeScheduler) {
     type,
     numTrainTimesteps: modelScheduler.num_train_timesteps ?? runtimeScheduler.numTrainTimesteps,
     shift: modelScheduler.shift ?? runtimeScheduler.shift,
+    predictionType: modelScheduler.prediction_type ?? runtimeScheduler.predictionType,
+    sigmaData: modelScheduler.sigma_data ?? runtimeScheduler.sigmaData,
   };
 }
 
@@ -95,6 +102,13 @@ export function initializeDiffusion(manifest, runtimeConfig) {
     }
     throw new Error('Diffusion manifest missing config.diffusion model contract.');
   }
+  const layout = modelConfig.layout;
+  if (layout && !SUPPORTED_DIFFUSION_RUNTIME_LAYOUTS.has(layout)) {
+    throw new Error(
+      `Diffusion layout "${layout}" is recognized in the manifest, but the GPU runtime is not implemented yet. ` +
+      'Supported runtime layouts: sd3, flux, sana.'
+    );
+  }
 
   const runtimeBase = mergeDiffusionConfig(DEFAULT_DIFFUSION_CONFIG, runtimeConfig?.inference?.diffusion);
   const runtime = {