feat(ui): LLM benchmark set say what it will actually run

flutter/analysis_options.yaml

@@ -4,6 +4,10 @@ analyzer:
   exclude:
     - lib/localizations/**
     - lib/protos/**
+    - build/**
+    - android/**
+    - ios/**
+    - windows/**
   language:
     strict-casts: true
 

flutter/lib/app_constants.dart

@@ -31,23 +31,48 @@ class BenchmarkId {
   static const imageClassificationV2 = 'image_classification_v2';
   static const imageClassificationOfflineV2 = 'image_classification_offline_v2';
   static const stableDiffusion = 'stable_diffusion';
-  static const llm = 'llm-1b';
-  static const llmInstruct = 'llm-1b-instruct';
+  static const llm1b = 'llm-1b';
+  static const llm1bInstruct = 'llm-1b-instruct';
+  static const llm3b = 'llm-3b';
+  static const llm3bInstruct = 'llm-3b-instruct';
+  static const llm8b = 'llm-8b';
+  static const llm8bInstruct = 'llm-8b-instruct';
 
-  // The sort order of this list will be used in the UI
+  static const llmIds = [
+    llm1b,
+    llm1bInstruct,
+    llm3b,
+    llm3bInstruct,
+    llm8b,
+    llm8bInstruct,
+  ];
+
+  // The sort order of this list will be used in the UI.
+  // Every task in tasks.pbtxt must appear here: BenchmarkStore sorts with
+  // indexOf, which returns -1 for anything missing and floats it to the top.
   static const allIds = [
     imageClassificationV2,
     objectDetection,
     imageSegmentationV2,
     naturalLanguageProcessing,
     superResolution,
     stableDiffusion,
-    llm,
-    llmInstruct,
+    llm1b,
+    llm1bInstruct,
+    llm3b,
+    llm3bInstruct,
+    llm8b,
+    llm8bInstruct,
     imageClassificationOfflineV2,
   ];
 }
 
+/// Ids of the `task_set` blocks in tasks.pbtxt.
+class BenchmarkSetId {
+  static const imageClassification = 'image_classification';
+  static const llm = 'llm';
+}
+
 class BackendId {
   static const tflite = 'libtflitebackend';
   static const pixel = 'libtflitepixelbackend';

flutter/lib/benchmark/benchmark.dart

@@ -188,6 +188,12 @@ class BenchmarkSet {
     applyOptionStateMap(optionState);
   }
 
+  BenchmarkSetInfo get info => BenchmarkSetInfo(config);
+
+  /// The benchmarks this set will actually run under the current options.
+  List<Benchmark> get activeBenchmarks =>
+      benchmarks.where((e) => e.isActive).toList();
+
   Iterable<BenchmarkOption> availableOptions() {
     var unhiddenOptions = optionSets
         .where((e) => !e.config.hidden)
@@ -277,6 +283,10 @@ class BenchmarkOptionSet {
     };
   }
 
+  /// True when the config allows exactly one option here, so the UI should
+  /// offer a single choice rather than independent checkboxes.
+  bool get isSingleChoice => config.maxSelected == 1;
+
   bool? getOption(String id) {
     return options[id]?.enabled;
   }

flutter/lib/benchmark/info.dart

@@ -27,9 +27,16 @@
 
   BenchmarkLocalizationInfo getLocalizedInfo(AppLocalizations stringResources) {
     switch (task.id) {
-      case (BenchmarkId.llm):
+      // Every LLM variant shares one description; the size and the eval it
+      // runs are already in the task's own name.
+      case (BenchmarkId.llm1b):
+      case (BenchmarkId.llm1bInstruct):
+      case (BenchmarkId.llm3b):
+      case (BenchmarkId.llm3bInstruct):
+      case (BenchmarkId.llm8b):
+      case (BenchmarkId.llm8bInstruct):
         return BenchmarkLocalizationInfo(
-          name: 'LLM',
+          name: task.name,
           detailsTitle: stringResources.benchInfoLlm,
           detailsContent: stringResources.benchInfoLlmDesc,
         );
@@ -91,3 +98,30 @@
   @override
   String toString() => 'Benchmark:${task.id}';
 }
+
+/// Name, icon and description of a whole task set.
+///
+/// A set is a frontend-only grouping, so it needs its own identity rather than
+/// borrowing one from whichever benchmark happens to sort first.
+class BenchmarkSetInfo {
+  final pb.TaskSet config;
+
+  BenchmarkSetInfo(this.config);
+
+  String get name => config.name;
+
+  Widget get icon => BenchmarkIcons.getSetDarkIcon(config.id);
+
+  /// Null when the set has no description of its own. Sets come from config,
+  /// so an unknown one is a gap to fall back from, not a crash.
+  String? localizedDetails(AppLocalizations stringResources) {
+    switch (config.id) {
+      case (BenchmarkSetId.llm):
+        return stringResources.benchInfoLlmDesc;
+      case (BenchmarkSetId.imageClassification):
+        return stringResources.benchInfoImageClassificationV2Desc;
+      default:
+        return null;
+    }
+  }
+}

flutter/lib/benchmark/state.dart

@@ -402,9 +402,30 @@ class BenchmarkState extends ChangeNotifier {
       return;
     }
     if (!benchmark.selectBackend(libName)) return;
-    // Persist only this explicit choice, merged into the stored map, so
-    // untouched benchmarks keep following future defaults and selections
-    // made under other task configs are preserved.
+    _persistBackendSelection({benchmark.id: libName});
+    notifyListeners();
+  }
+
+  /// Applies one backend to every benchmark in the set that offers it.
+  /// Benchmarks without that backend keep the one they have.
+  void benchmarkSetBackendForSet(BenchmarkSet benchmarkSet, String libName) {
+    if (state == BenchmarkStateEnum.running ||
+        state == BenchmarkStateEnum.aborting) {
+      return;
+    }
+    final applied = <String, String>{};
+    for (final benchmark in benchmarkSet.benchmarks) {
+      if (benchmark.selectBackend(libName)) applied[benchmark.id] = libName;
+    }
+    if (applied.isEmpty) return;
+    _persistBackendSelection(applied);
+    notifyListeners();
+  }
+
+  /// Persist only explicit choices, merged into the stored map, so untouched
+  /// benchmarks keep following future defaults and selections made under other
+  /// task configs are preserved.
+  void _persistBackendSelection(Map<String, String> choices) {
     Map<String, dynamic> stored = {};
     if (_store.backendSelection.isNotEmpty) {
       try {
@@ -413,12 +434,11 @@ class BenchmarkState extends ChangeNotifier {
         print('Backend selection parse fail: $e');
       }
     }
-    stored[benchmark.id] = libName;
+    stored.addAll(choices);
     _store.backendSelection = jsonEncode(stored);
     // The resource map is built for the previously selected backends;
-    // rebuild it so resource validation sees the new backend's files.
+    // rebuild it so resource validation sees the new backends' files.
     deferredLoadResources();
-    notifyListeners();
   }
 
   void benchmarkSetOption(

flutter/lib/l10n/app_en.arb

@@ -35,7 +35,18 @@
   "mainScreenGo": "GO",
   "mainScreenMeasureTitle": "Measure your device performance for:",
   "mainScreenBenchmarkSelected": "Benchmarks (<selected>/<total> selected)",
-  "mainScreenOptionsSelected": "<active>/<total> options selected",
+  "mainScreenSetRunsCount": "Runs <count> of <total> benchmarks",
+  "mainScreenSetRunsNothing": "Runs nothing",
+  "mainScreenSetNothingSelected": "Nothing selected — this set will not run",
+  "mainScreenSetBackends": "Backends",
+  "mainScreenSetFilesMissing": "Files missing",
+  "mainScreenSetBackendForAll": "Apply to all",
+  "mainScreenSetBackendMixed": "Mixed",
+  "mainScreenBackendLabel": "Backend",
+  "mainScreenDelegateLabel": "Delegate",
+  "mainScreenOptionRuleOneOrMore": "select one or more",
+  "mainScreenOptionRulePickOne": "pick one",
+  "mainScreenBenchmarkNotSelected": "not selected",
 
   "progressPerformance": "Performance",
   "progressAccuracy": "Accuracy",
@@ -55,7 +66,8 @@
   "resultsNotAvailable": "N/A",
   "resultsBenchmarkAborted": "Benchmark aborted!",
   "resultsBatchSize": "Batches: <batchSize>",
-  "resultsBenchmarksInSet": "<count> benchmarks in this set",
+  "resultsSetRanCount": "<count> of <total> benchmarks ran",
+  "resultsBenchmarkNotRun": "Not run",
 
   "shareButtonMLCommons": "Share with MLCommons",
   "shareButtonOther": "Share with others",

flutter/lib/l10n/app_zh.arb

@@ -35,7 +35,18 @@
   "mainScreenGo": "开始",
   "mainScreenMeasureTitle": "测量您的设备性能：",
   "mainScreenBenchmarkSelected": "基准测试（已选择 <selected>/<total>）",
-  "mainScreenOptionsSelected": "已选择 <active>/<total> 个选项",
+  "mainScreenSetRunsCount": "运行 <total> 项中的 <count> 项基准测试",
+  "mainScreenSetRunsNothing": "不运行任何基准测试",
+  "mainScreenSetNothingSelected": "未选择任何选项 — 此组不会运行",
+  "mainScreenSetBackends": "后端",
+  "mainScreenSetFilesMissing": "缺少文件",
+  "mainScreenSetBackendForAll": "应用于全部",
+  "mainScreenSetBackendMixed": "混合",
+  "mainScreenBackendLabel": "后端",
+  "mainScreenDelegateLabel": "委托",
+  "mainScreenOptionRuleOneOrMore": "选择一项或多项",
+  "mainScreenOptionRulePickOne": "选择一项",
+  "mainScreenBenchmarkNotSelected": "未选择",
 
   "progressPerformance": "性能",
   "progressAccuracy": "准确度",
@@ -55,7 +66,8 @@
   "resultsNotAvailable": "N/A",
   "resultsBenchmarkAborted": "基准测试已中止！",
   "resultsBatchSize": "批次：<batchSize>",
-  "resultsBenchmarksInSet": "此组包含 <count> 个基准测试",
+  "resultsSetRanCount": "已运行 <total> 项中的 <count> 项基准测试",
+  "resultsBenchmarkNotRun": "未运行",
 
   "shareButtonMLCommons": "分享给 MLCommons",
   "shareButtonOther": "分享给他人",

flutter/lib/ui/app_styles.dart

@@ -40,6 +40,21 @@
   static const progressCircle = _blue4;
   static const goCircle = _green;
   static const warningIcon = _yellow;
+
+  // Benchmark config / result cards.
+  static const cardBorder = Color(0xFFE2E8F0);
+  static const cardDivider = Color(0xFFF1F5F9);
+  static const cardFooter = Color(0xFFFAFBFC);
+  static const panelBackground = Color(0xFFF4F7FA);
+  static const panelDivider = Color(0xFFE6ECF1);
+  static const setIconBackground = Color(0xFFEAF4FB);
+  static const chipBorder = Color(0xFFC3CED9);
+  static const bodyText = Color(0xFF0F172A);
+  static const mutedText = Color(0xFF475569);
+  static const subtleText = Color(0xFF64748B);
+  static const warningText = Color(0xFF7A5800);
+  static const warningBackground = Color(0xFFFDF6E0);
+  static const warningBorder = Color(0xFFE8C95C);
 }
 
 class AppGradients {