Add support for performance hints when loading on-device model

This allows passing a performance hint to the backend when loading the
on-device model. This can be used to tune various params in the backend
or use a smaller model if requested.

This is controlled from the optimization guide side by a new feature
param "compatible_low_tier_on_device_performance_classes", which
allows us to choose a set of performance classes that will use the
kFastestInference performance hint.

As part of this change, all the feature params related to model
performance were moved under a new base::Feature to allow them to be
adjusted independent of the other on-device related params.

Bug: 379723772
Change-Id: I28db8de699a8af026ee2ee1cfa08136342860e13
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/6089104
Reviewed-by: Matthew Denton <mpdenton@chromium.org>
Commit-Queue: Clark DuVall <cduvall@chromium.org>
Reviewed-by: Sophie Chang <sophiechang@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1395676}

chrome/browser/about_flags.cc

@@ -6645,7 +6645,7 @@ const FeatureEntry kFeatureEntries[] = {
      flag_descriptions::kOptimizationGuideOnDeviceModelName,
      flag_descriptions::kOptimizationGuideOnDeviceModelDescription, kOsDesktop,
      FEATURE_WITH_PARAMS_VALUE_TYPE(
-         optimization_guide::features::kOptimizationGuideOnDeviceModel,
+         optimization_guide::features::kOnDeviceModelPerformanceParams,
          kOptimizationGuideOnDeviceModelVariations,
          "OptimizationGuideOnDeviceModel")},
 

chrome/browser/optimization_guide/model_execution/model_execution_browsertest.cc

@@ -690,7 +690,8 @@ class OnDeviceModelExecutionEnabledBrowserTest
     scoped_feature_list_.InitWithFeaturesAndParameters(
         {{features::kOptimizationGuideModelExecution, {}},
          {features::kModelQualityLogging, {}},
-         {features::kOptimizationGuideOnDeviceModel,
+         {features::kOptimizationGuideOnDeviceModel, {}},
+         {features::kOnDeviceModelPerformanceParams,
           {{"compatible_on_device_performance_classes", "*"}}}},
         {});
   }

chrome/browser/resources/on_device_internals/BUILD.gn

@@ -16,6 +16,7 @@ build_webui("build") {
   mojo_files = [
     "$root_gen_dir/chrome/browser/ui/webui/on_device_internals/on_device_internals_page.mojom-webui.ts",
     "$root_gen_dir/services/on_device_model/public/mojom/on_device_model.mojom-webui.ts",
+    "$root_gen_dir/services/on_device_model/public/mojom/on_device_model_service.mojom-webui.ts",
   ]
 
   static_files = [ "on_device_internals.html" ]

chrome/browser/resources/on_device_internals/tools.html

@@ -1,4 +1,4 @@
-<style include="cr-shared-style cr-hidden-style">
+<style include="cr-shared-style cr-hidden-style md-select">
   :host {
     display: block;
     margin: auto;
@@ -110,24 +110,40 @@
   cr-expand-button:hover {
     background-color: var(--cr-hover-background-color);
   }
+
+  .model-options {
+    display: flex;
+    flex-direction: row;
+  }
+
+  cr-checkbox {
+    margin-left: 20px;
+    --cr-checkbox-label-padding-start: 10px;
+  }
 </style>
 <div class="performance-class">
   Device performance class: <strong>[[performanceClassText_]]</strong>
 </div>
 <cr-input id="modelInput" label="Model directory" placeholder="/tmp/model"
     disabled="[[isLoading_(loadModelStart_)]]"
-    on-change="onModelSelected_" error-message="[[error_]]"
-    invalid="[[error_.length]]" autofocus>
+    error-message="[[error_]]" invalid="[[error_.length]]" autofocus>
   <cr-button slot="suffix" disabled="[[isLoading_(loadModelStart_)]]"
       on-click="onLoadClick_">
     Load
   </cr-button>
 </cr-input>
-<cr-checkbox slot="suffix" checked="{{enableImageInput_}}">
-  Enable images
-</cr-checkbox>
+<div class="model-options">
+  <select id="performanceHintSelect" class="md-select"
+      value="[[performanceHint_]]" on-change="onPerformanceHintChange_">
+    <option value="kHighestQuality">Highest Quality</option>
+    <option value="kFastestInference">Fastest Inference</option>
+  </select>
+  <cr-checkbox slot="suffix" checked="{{enableImageInput_}}">
+    Enable images
+  </cr-checkbox>
+</div>
 <div class="model-text">
-  [[getModelText_(modelPath_, loadModelDuration_)]]
+  [[getModelText_(modelPath_, loadModelDuration_, loadedPerformanceHint_)]]
   <div class="throbber" hidden$="[[!isLoading_(loadModelStart_)]]"></div>
 </div>
 

chrome/browser/resources/on_device_internals/tools.ts

@@ -10,6 +10,7 @@ import '//resources/cr_elements/cr_hidden_style.css.js';
 import '//resources/cr_elements/cr_input/cr_input.js';
 import '//resources/cr_elements/cr_shared_vars.css.js';
 import '//resources/cr_elements/cr_textarea/cr_textarea.js';
+import '//resources/cr_elements/md_select.css.js';
 
 import type {CrInputElement} from '//resources/cr_elements/cr_input/cr_input.js';
 import {PolymerElement} from '//resources/polymer/v3_0/polymer/polymer_bundled.min.js';
@@ -17,6 +18,7 @@ import {PolymerElement} from '//resources/polymer/v3_0/polymer/polymer_bundled.m
 import {BrowserProxy} from './browser_proxy.js';
 import type {InputPiece, ResponseChunk, ResponseSummary} from './on_device_model.mojom-webui.js';
 import {LoadModelResult, OnDeviceModelRemote, PerformanceClass, SessionRemote, StreamingResponderCallbackRouter, Token} from './on_device_model.mojom-webui.js';
+import {ModelPerformanceHint} from './on_device_model_service.mojom-webui.js';
 import {getTemplate} from './tools.html.js';
 
 interface Response {
@@ -34,6 +36,7 @@ interface OnDeviceInternalsToolsElement {
     textInput: CrInputElement,
     imageInput: HTMLInputElement,
     topKInput: CrInputElement,
+    performanceHintSelect: HTMLSelectElement,
   };
 }
 
@@ -131,6 +134,8 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
         type: Object,
         value: null,
       },
+      performanceHint_: String,
+      loadedPerformanceHint_: Number,
     };
   }
 
@@ -159,6 +164,8 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
   private topK_: number = 1;
   private imageFile_: File|null = null;
   private enableImageInput_: boolean = false;
+  private performanceHint_: string = 'kHighestQuality';
+  private loadedPerformanceHint_: ModelPerformanceHint|null;
 
   private proxy_: BrowserProxy = BrowserProxy.getInstance();
   private responseRouter_: StreamingResponderCallbackRouter =
@@ -199,6 +206,10 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
     this.$.imageInput.value = '';
   }
 
+  private onPerformanceHintChange_() {
+    this.performanceHint_ = this.$.performanceHintSelect.value;
+  }
+
   private onServiceCrashed_() {
     if (this.currentResponse_) {
       this.currentResponse_.error = true;
@@ -233,6 +244,8 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
     this.baseModel_ = null;
     this.model_ = null;
     this.loadModelStart_ = new Date().getTime();
+    const performanceHint = ModelPerformanceHint[(
+        this.performanceHint_ as keyof typeof ModelPerformanceHint)];
     const modelPath = this.$.modelInput.value;
     // <if expr="is_win">
     // Windows file paths are std::wstring, so use Array<Number>.
@@ -244,7 +257,8 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
     const baseModel = new OnDeviceModelRemote();
     let newModel = new OnDeviceModelRemote();
     let {result} = await this.proxy_.handler.loadModel(
-        {path: processedPath}, baseModel.$.bindNewPipeAndPassReceiver());
+        {path: processedPath}, performanceHint,
+        baseModel.$.bindNewPipeAndPassReceiver());
     if (result === LoadModelResult.kSuccess && this.enableImageInput_) {
       result = (await baseModel.loadAdaptation(
                     {
@@ -272,6 +286,7 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
       });
       this.startNewSession_();
       this.modelPath_ = modelPath;
+      this.loadedPerformanceHint_ = performanceHint;
     }
   }
 
@@ -419,9 +434,13 @@ class OnDeviceInternalsToolsElement extends PolymerElement {
       return '';
     }
     let text = 'Model loaded from ' + this.modelPath_ + ' in ' +
-        this.loadModelDuration_ + 'ms';
+        this.loadModelDuration_ + 'ms ';
     if (this.imagesEnabled_()) {
-      text += ' [images enabled]';
+      text += '[images enabled]';
+    }
+    if (this.loadedPerformanceHint_ ===
+        ModelPerformanceHint.kFastestInference) {
+      text += '[fastest inference]';
     }
     return text;
   }

chrome/browser/ui/webui/on_device_internals/on_device_internals_page.mojom

@@ -8,6 +8,7 @@ import "mojo/public/mojom/base/big_buffer.mojom";
 import "mojo/public/mojom/base/file_path.mojom";
 import "mojo/public/mojom/base/time.mojom";
 import "services/on_device_model/public/mojom/on_device_model.mojom";
+import "services/on_device_model/public/mojom/on_device_model_service.mojom";
 import "skia/public/mojom/bitmap.mojom";
 
 // Struct containing data to be displayed on on-device-internals page.
@@ -43,6 +44,7 @@ interface OnDeviceInternalsPageHandler {
   // Binds a new OnDeviceModel interface if possible using model assets loaded
   // from within `model_path`.
   LoadModel(mojo_base.mojom.FilePath model_path,
+            on_device_model.mojom.ModelPerformanceHint performance_hint,
             pending_receiver<on_device_model.mojom.OnDeviceModel> model) =>
       (on_device_model.mojom.LoadModelResult result);
 

chrome/browser/ui/webui/on_device_internals/on_device_internals_page_handler.cc

@@ -61,6 +61,7 @@ OnDeviceInternalsPageHandler::~OnDeviceInternalsPageHandler() {
 
 void OnDeviceInternalsPageHandler::LoadModel(
     const base::FilePath& model_path,
+    ml::ModelPerformanceHint performance_hint,
     mojo::PendingReceiver<on_device_model::mojom::OnDeviceModel> model,
     LoadModelCallback callback) {
 #if BUILDFLAG(USE_CHROMEOS_MODEL_SERVICE)
@@ -81,7 +82,7 @@ void OnDeviceInternalsPageHandler::LoadModel(
       base::BindOnce(&LoadModelAssets, model_path),
       base::BindOnce(&OnDeviceInternalsPageHandler::OnModelAssetsLoaded,
                      weak_ptr_factory_.GetWeakPtr(), std::move(model),
-                     std::move(callback)));
+                     std::move(callback), performance_hint));
 #endif
 }
 
@@ -109,10 +110,12 @@ OnDeviceInternalsPageHandler::GetService() {
 void OnDeviceInternalsPageHandler::OnModelAssetsLoaded(
     mojo::PendingReceiver<on_device_model::mojom::OnDeviceModel> model,
     LoadModelCallback callback,
+    ml::ModelPerformanceHint performance_hint,
     on_device_model::ModelAssets assets) {
   auto params = on_device_model::mojom::LoadModelParams::New();
   params->assets = std::move(assets);
   params->max_tokens = 4096;
+  params->performance_hint = performance_hint;
   GetService().LoadModel(std::move(params), std::move(model),
                          std::move(callback));
 }

chrome/browser/ui/webui/on_device_internals/on_device_internals_page_handler.h

@@ -43,12 +43,14 @@ class OnDeviceInternalsPageHandler : public mojom::OnDeviceInternalsPageHandler,
   void OnModelAssetsLoaded(
       mojo::PendingReceiver<on_device_model::mojom::OnDeviceModel> model,
       LoadModelCallback callback,
+      ml::ModelPerformanceHint performance_hint,
       on_device_model::ModelAssets assets);
 #endif
 
   // mojom::OnDeviceInternalsPageHandler:
   void LoadModel(
       const base::FilePath& model_path,
+      ml::ModelPerformanceHint performance_hint,
       mojo::PendingReceiver<on_device_model::mojom::OnDeviceModel> model,
       LoadModelCallback callback) override;
   void GetEstimatedPerformanceClass(

components/optimization_guide/core/model_execution/on_device_model_component.cc

@@ -138,6 +138,13 @@ OnDeviceModelComponentStateManager::GetRegistrationCriteria() {
   return registration_criteria_.get();
 }
 
+bool OnDeviceModelComponentStateManager::IsLowTierDevice() const {
+  DCHECK_CALLED_ON_VALID_SEQUENCE(sequence_checker_);
+  return IsPerformanceClassCompatible(
+      features::kLowTierPerformanceClassListForOnDeviceModel.Get(),
+      PerformanceClassFromPref(*local_state_));
+}
+
 void OnDeviceModelComponentStateManager::OnDeviceEligibleFeatureUsed(
     ModelBasedCapabilityKey feature) {
   DCHECK_CALLED_ON_VALID_SEQUENCE(sequence_checker_);

components/optimization_guide/core/model_execution/on_device_model_component.h

@@ -179,6 +179,9 @@ class OnDeviceModelComponentStateManager
   // registration has been computed yet.
   const RegistrationCriteria* GetRegistrationCriteria();
 
+  // Returns true if this is determined to be a low tier device.
+  bool IsLowTierDevice() const;
+
   base::WeakPtr<OnDeviceModelComponentStateManager> GetWeakPtr() {
     return weak_ptr_factory_.GetWeakPtr();
   }

components/optimization_guide/core/model_execution/on_device_model_component_unittest.cc

@@ -53,7 +53,8 @@ class OnDeviceModelComponentTest : public testing::Test {
 
     feature_list_.InitWithFeaturesAndParameters(
         {{features::kOptimizationGuideModelExecution, {}},
-         {features::kOptimizationGuideOnDeviceModel,
+         {features::kOptimizationGuideOnDeviceModel, {}},
+         {features::kOnDeviceModelPerformanceParams,
           {{"compatible_on_device_performance_classes", "3,4,5,6"}}}},
         /*disabled_features=*/{});
   }

components/optimization_guide/core/model_execution/on_device_model_service_controller.cc

@@ -279,6 +279,10 @@ void OnDeviceModelServiceController::OnModelAssetsLoaded(
   // TODO(crbug.com/302402959): Choose max_tokens based on device.
   params->max_tokens = features::GetOnDeviceModelMaxTokens();
   params->adaptation_ranks = features::GetOnDeviceModelAllowedAdaptationRanks();
+  if (on_device_component_state_manager_ &&
+      on_device_component_state_manager_->IsLowTierDevice()) {
+    params->performance_hint = ml::ModelPerformanceHint::kFastestInference;
+  }
   service_client_.Get()->LoadModel(
       std::move(params), std::move(model),
       base::DoNothingAs<void(on_device_model::mojom::LoadModelResult)>());

components/optimization_guide/core/model_execution/on_device_model_service_controller_unittest.cc

@@ -213,6 +213,9 @@ class OnDeviceModelServiceControllerTest : public testing::Test {
            {"on_device_model_disable_crash_count", "3"},
            {"on_device_model_crash_backoff_base_time", "1m"},
            {"on_device_model_max_crash_backoff_time", "1h"}}},
+         {features::kOnDeviceModelPerformanceParams,
+          {{"compatible_on_device_performance_classes", "*"},
+           {"compatible_low_tier_on_device_performance_classes", "3"}}},
          {features::kTextSafetyClassifier, {}},
          {features::kOnDeviceModelValidation,
           {{"on_device_model_validation_delay", "0"}}}},
@@ -3851,4 +3854,17 @@ TEST_F(OnDeviceModelServiceControllerTest, LoggingModeAlwaysDisable) {
   EXPECT_EQ(0u, test_uploader.uploaded_logs().size());
 }
 
+TEST_F(OnDeviceModelServiceControllerTest, SendsPerformanceHint) {
+  // Low performance class should use fastest inference.
+  pref_service_.SetInteger(
+      model_execution::prefs::localstate::kOnDevicePerformanceClass,
+      base::to_underlying(OnDeviceModelPerformanceClass::kLow));
+  Initialize(standard_assets_);
+  auto session = CreateSession();
+  session->ExecuteModel(PageUrlRequest("foo"),
+                        response_.GetStreamingCallback());
+  ASSERT_TRUE(response_.GetFinalStatus());
+  EXPECT_EQ(*response_.value(), "Fastest inference\nInput: execute:foo\n");
+}
+
 }  // namespace optimization_guide

components/optimization_guide/core/optimization_guide_features.cc

@@ -202,10 +202,19 @@ BASE_FEATURE(kAiSettingsPageEnterpriseDisabledUi,
              "AiSettingsPageEnterpriseDisabledUi",
              base::FEATURE_DISABLED_BY_DEFAULT);
 
+BASE_FEATURE(kOnDeviceModelPerformanceParams,
+             "OnDeviceModelPerformanceParams",
+             base::FEATURE_ENABLED_BY_DEFAULT);
+
 const base::FeatureParam<std::string> kPerformanceClassListForOnDeviceModel{
-    &kOptimizationGuideOnDeviceModel,
+    &kOnDeviceModelPerformanceParams,
     "compatible_on_device_performance_classes", "5,6"};
 
+const base::FeatureParam<std::string>
+    kLowTierPerformanceClassListForOnDeviceModel{
+        &kOnDeviceModelPerformanceParams,
+        "compatible_low_tier_on_device_performance_classes", ""};
+
 BASE_FEATURE(kOptimizationGuideIconView,
              "OptimizationGuideIconView",
              base::FEATURE_DISABLED_BY_DEFAULT);

components/optimization_guide/core/optimization_guide_features.h

@@ -90,6 +90,11 @@ BASE_DECLARE_FEATURE(kPrivacyGuideAiSettings);
 COMPONENT_EXPORT(OPTIMIZATION_GUIDE_FEATURES)
 extern const base::FeatureParam<bool> kShowAiSettingsForTesting;
 
+// Allows setting feature params for model download configuration, such as
+// minimum performance class for download.
+COMPONENT_EXPORT(OPTIMIZATION_GUIDE_FEATURES)
+BASE_DECLARE_FEATURE(kOnDeviceModelPerformanceParams);
+
 // Comma-separated list of performance classes (e.g. "3,4,5") that should
 // download the base model. Use "*" if there is no performance class
 // requirement.
@@ -97,6 +102,13 @@ COMPONENT_EXPORT(OPTIMIZATION_GUIDE_FEATURES)
 extern const base::FeatureParam<std::string>
     kPerformanceClassListForOnDeviceModel;
 
+// Comma-separated list of performance classes that should use a smaller model
+// if available. This should be a subset of
+// kPerformanceClassListForOnDeviceModel.
+COMPONENT_EXPORT(OPTIMIZATION_GUIDE_FEATURES)
+extern const base::FeatureParam<std::string>
+    kLowTierPerformanceClassListForOnDeviceModel;
+
 COMPONENT_EXPORT(OPTIMIZATION_GUIDE_FEATURES)
 BASE_DECLARE_FEATURE(kOptimizationGuideIconView);
 

services/on_device_model/fake/fake_chrome_ml_api.cc

@@ -65,7 +65,8 @@ bool QueryGPUAdapter(void (*adapter_callback_fn)(WGPUAdapter adapter,
 }
 
 struct FakeModelInstance {
-  ModelBackendType backend_type_;
+  ml::ModelBackendType backend_type_;
+  ml::ModelPerformanceHint performance_hint;
   std::string model_data_;
 };
 
@@ -97,8 +98,10 @@ std::string ReadFile(PlatformFile api_file) {
 ChromeMLModel SessionCreateModel(const ChromeMLModelDescriptor* descriptor,
                                  uintptr_t context,
                                  ChromeMLScheduleFn schedule) {
-  return reinterpret_cast<ChromeMLModel>(
-      new FakeModelInstance{.backend_type_ = descriptor->backend_type});
+  return reinterpret_cast<ChromeMLModel>(new FakeModelInstance{
+      .backend_type_ = descriptor->backend_type,
+      .performance_hint = descriptor->performance_hint,
+  });
 }
 
 void DestroyModel(ChromeMLModel model) {
@@ -121,11 +124,11 @@ ChromeMLSession CreateSession(ChromeMLModel model,
   if (descriptor) {
     instance->enable_image_input = descriptor->enable_image_input;
     if (descriptor->model_data) {
-      if (model_instance->backend_type_ == ModelBackendType::kGpuBackend) {
+      if (model_instance->backend_type_ == ml::ModelBackendType::kGpuBackend) {
         instance->adaptation_data_ =
             ReadFile(descriptor->model_data->weights_file);
       } else if (model_instance->backend_type_ ==
-                 ModelBackendType::kApuBackend) {
+                 ml::ModelBackendType::kApuBackend) {
         base::ReadFileToString(
             base::FilePath::FromUTF8Unsafe(descriptor->model_data->model_path),
             &instance->adaptation_data_);
@@ -206,6 +209,10 @@ bool SessionExecuteModel(ChromeMLSession session,
         output_fn(&output);
       };
 
+  if (reinterpret_cast<FakeModelInstance*>(model)->performance_hint ==
+      ml::ModelPerformanceHint::kFastestInference) {
+    OutputChunk("Fastest inference\n");
+  }
   if (!instance->adaptation_data_.empty()) {
     OutputChunk("Adaptation: " + instance->adaptation_data_ + "\n");
   }

services/on_device_model/ml/chrome_ml_api.h

@@ -15,6 +15,9 @@
 
 // This header defines the public interface to the ChromeML shared library.
 
+// TODO: crbug.com/379723772 - Remove this when internal code migrates.
+using ::ml::ModelBackendType;
+
 extern "C" {
 
 // A function used to handle fatal errors.
@@ -43,15 +46,6 @@ using ChromeMLTSModel = uintptr_t;
 // Opaque handle to a video-frame-specific ML inference engine.
 using ChromeMLInferenceEngine = uintptr_t;
 
-// Type of the backend to run the model.
-enum ModelBackendType {
-  // The default WebGPU backend.
-  kGpuBackend = 0,
-  // The APU accelerator backend. Only available on devices with APU, and need
-  // special APU model files.
-  kApuBackend = 1,
-};
-
 // A contiguous byte span.
 struct ChromeMLByteSpan {
   uint8_t* data;
@@ -79,7 +73,7 @@ struct ChromeMLModelData {
 // Describes a model to use with ChromeML.
 struct ChromeMLModelDescriptor {
   // The backend to run this model.
-  ModelBackendType backend_type;
+  ml::ModelBackendType backend_type;
 
   // The model data to use.
   const ChromeMLModelData* model_data;
@@ -105,6 +99,8 @@ struct ChromeMLModelDescriptor {
   bool enable_host_mapped_pointer;
   bool use_low_power;
   bool allow_fp16;
+
+  ml::ModelPerformanceHint performance_hint;
 };
 
 // Describes an adaptation for a model.

services/on_device_model/ml/chrome_ml_types.h

@@ -28,6 +28,21 @@ enum class Token {
 // current library version.
 using InputPiece = std::variant<Token, std::string, SkBitmap, bool>;
 
+// Options for specifying the performance characteristics of the model to load.
+enum class ModelPerformanceHint {
+  kHighestQuality,
+  kFastestInference,
+};
+
+// Type of the backend to run the model.
+enum ModelBackendType {
+  // The default WebGPU backend.
+  kGpuBackend = 0,
+  // The APU accelerator backend. Only available on devices with APU, and need
+  // special APU model files.
+  kApuBackend = 1,
+};
+
 }  // namespace ml
 
 #endif  // SERVICES_ON_DEVICE_MODEL_ML_CHROME_ML_TYPES_H_

services/on_device_model/ml/chrome_ml_types_traits.cc

@@ -98,4 +98,62 @@ bool UnionTraits<on_device_model::mojom::InputPieceDataView, ml::InputPiece>::
   return false;
 }
 
+// static
+on_device_model::mojom::ModelBackendType
+EnumTraits<on_device_model::mojom::ModelBackendType,
+           ml::ModelBackendType>::ToMojom(ml::ModelBackendType input) {
+  switch (input) {
+    case ml::ModelBackendType::kGpuBackend:
+      return on_device_model::mojom::ModelBackendType::kGpu;
+    case ml::ModelBackendType::kApuBackend:
+      return on_device_model::mojom::ModelBackendType::kApu;
+  }
+  NOTREACHED();
+}
+
+// static
+bool EnumTraits<on_device_model::mojom::ModelBackendType,
+                ml::ModelBackendType>::
+    FromMojom(on_device_model::mojom::ModelBackendType input,
+              ml::ModelBackendType* output) {
+  switch (input) {
+    case on_device_model::mojom::ModelBackendType::kGpu:
+      *output = ml::ModelBackendType::kGpuBackend;
+      return true;
+    case on_device_model::mojom::ModelBackendType::kApu:
+      *output = ml::ModelBackendType::kApuBackend;
+      return true;
+  }
+  return false;
+}
+
+// static
+on_device_model::mojom::ModelPerformanceHint
+EnumTraits<on_device_model::mojom::ModelPerformanceHint,
+           ml::ModelPerformanceHint>::ToMojom(ml::ModelPerformanceHint input) {
+  switch (input) {
+    case ml::ModelPerformanceHint::kHighestQuality:
+      return on_device_model::mojom::ModelPerformanceHint::kHighestQuality;
+    case ml::ModelPerformanceHint::kFastestInference:
+      return on_device_model::mojom::ModelPerformanceHint::kFastestInference;
+  }
+  NOTREACHED();
+}
+
+// static
+bool EnumTraits<on_device_model::mojom::ModelPerformanceHint,
+                ml::ModelPerformanceHint>::
+    FromMojom(on_device_model::mojom::ModelPerformanceHint input,
+              ml::ModelPerformanceHint* output) {
+  switch (input) {
+    case on_device_model::mojom::ModelPerformanceHint::kHighestQuality:
+      *output = ml::ModelPerformanceHint::kHighestQuality;
+      return true;
+    case on_device_model::mojom::ModelPerformanceHint::kFastestInference:
+      *output = ml::ModelPerformanceHint::kFastestInference;
+      return true;
+  }
+  return false;
+}
+
 }  // namespace mojo

services/on_device_model/ml/chrome_ml_types_traits.h

@@ -43,6 +43,24 @@ struct UnionTraits<on_device_model::mojom::InputPieceDataView, ml::InputPiece> {
                    ml::InputPiece* out);
 };
 
+template <>
+struct EnumTraits<on_device_model::mojom::ModelBackendType,
+                  ml::ModelBackendType> {
+  static on_device_model::mojom::ModelBackendType ToMojom(
+      ml::ModelBackendType input);
+  static bool FromMojom(on_device_model::mojom::ModelBackendType input,
+                        ml::ModelBackendType* output);
+};
+
+template <>
+struct EnumTraits<on_device_model::mojom::ModelPerformanceHint,
+                  ml::ModelPerformanceHint> {
+  static on_device_model::mojom::ModelPerformanceHint ToMojom(
+      ml::ModelPerformanceHint input);
+  static bool FromMojom(on_device_model::mojom::ModelPerformanceHint input,
+                        ml::ModelPerformanceHint* output);
+};
+
 }  // namespace mojo
 
 #endif  // SERVICES_ON_DEVICE_MODEL_ML_CHROME_ML_TYPES_TRAITS_H_

services/on_device_model/ml/gpu_blocklist.cc

@@ -16,7 +16,7 @@ namespace ml {
 namespace {
 
 const base::FeatureParam<std::string> kGpuBlockList{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_model_gpu_block_list",
     // These devices are nearly always crashing or have very low performance.
     "8086:412|8086:a16|8086:41e|8086:416|8086:402|8086:166|8086:1616|8086:22b1|"

services/on_device_model/ml/on_device_model_executor.cc

@@ -105,18 +105,6 @@ uint32_t GetTopK(std::optional<uint32_t> top_k) {
                   std::max(1u, top_k.value_or(1)));
 }
 
-std::optional<ModelBackendType> ModelBackendTypeFromMojom(
-    on_device_model::mojom::ModelBackendType backend) {
-  switch (backend) {
-    case on_device_model::mojom::ModelBackendType::kGpu:
-      return ModelBackendType::kGpuBackend;
-    case on_device_model::mojom::ModelBackendType::kApu:
-      return ModelBackendType::kApuBackend;
-    default:
-      return std::nullopt;
-  }
-}
-
 }  // namespace
 
 // Handles sending and canceling responses.
@@ -466,24 +454,17 @@ LoadModelResult OnDeviceModelExecutor::Init(
 
   max_tokens_ = std::max(params->max_tokens, kReserveTokensForSafety);
 
-  std::optional<ModelBackendType> backend_type =
-      ModelBackendTypeFromMojom(params->backend_type);
-  if (!backend_type.has_value()) {
-    LOG(ERROR) << "Failed to parse model backend type";
-    return LoadModelResult::kFailedToLoadLibrary;
-  }
-
   ChromeMLModelData data;
   std::string weights_path_str = assets.weights_path.AsUTF8Unsafe();
   std::string sp_model_path_str = assets.sp_model_path.AsUTF8Unsafe();
-  if (*backend_type == ModelBackendType::kGpuBackend) {
+  if (params->backend_type == ml::ModelBackendType::kGpuBackend) {
     data.weights_file = assets.weights.TakePlatformFile();
   } else {
     data.model_path = weights_path_str.data();
     data.sentencepiece_model_path = sp_model_path_str.data();
   }
   ChromeMLModelDescriptor descriptor = {
-      .backend_type = *backend_type,
+      .backend_type = params->backend_type,
       .model_data = &data,
       .max_tokens = max_tokens_,
       .temperature = 0.0f,
@@ -494,6 +475,7 @@ LoadModelResult OnDeviceModelExecutor::Init(
       .enable_host_mapped_pointer = kEnableHostMappedPointer.Get(),
       .use_low_power = kUseLowPower.Get(),
       .allow_fp16 = kAllowFp16.Get(),
+      .performance_hint = params->performance_hint,
   };
   model_ = chrome_ml_->api().SessionCreateModel(
       &descriptor, reinterpret_cast<uintptr_t>(this),

services/on_device_model/ml/performance_class.cc

@@ -18,30 +18,30 @@ constexpr uint64_t kBytesPerMb = 1024 * 1024;
 
 // The threshold for GPU RAM below which the device is considered VeryLow.
 const base::FeatureParam<int> kLowRAMThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_low_ram_threshold_mb", 3000};
 // RAM threshold necessary to be considered High or better.
 const base::FeatureParam<int> kHighRAMThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_high_ram_threshold_mb", 5500};
 
 // Output threshold to be considered Low or better.
 const base::FeatureParam<int> kLowOutputThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_low_output_threshold", 5};
 
 // Input speed min thresholds or each device class.
 const base::FeatureParam<int> kLowThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_low_threshold", 50};
 const base::FeatureParam<int> kMediumThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_medium_threshold", 75};
 const base::FeatureParam<int> kHighThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_high_threshold", 150};
 const base::FeatureParam<int> kVeryHighThreshold{
-    &optimization_guide::features::kOptimizationGuideOnDeviceModel,
+    &optimization_guide::features::kOnDeviceModelPerformanceParams,
     "on_device_very_high_threshold", 500};
 
 // These values are persisted to logs. Entries should not be renumbered and

services/on_device_model/on_device_model_service_unittest.cc

@@ -81,11 +81,14 @@ class OnDeviceModelServiceTest : public testing::Test {
   mojo::Remote<mojom::OnDeviceModelService>& service() { return service_; }
 
   mojo::Remote<mojom::OnDeviceModel> LoadModel(
-      mojom::ModelBackendType backend_type = mojom::ModelBackendType::kGpu) {
+      ml::ModelBackendType backend_type = ml::ModelBackendType::kGpuBackend,
+      ml::ModelPerformanceHint performance_hint =
+          ml::ModelPerformanceHint::kHighestQuality) {
     base::RunLoop run_loop;
     mojo::Remote<mojom::OnDeviceModel> remote;
     auto params = mojom::LoadModelParams::New();
     params->backend_type = backend_type;
+    params->performance_hint = performance_hint;
     params->max_tokens = 8000;
     service()->LoadModel(
         std::move(params), remote.BindNewPipeAndPassReceiver(),
@@ -452,7 +455,7 @@ TEST_F(OnDeviceModelServiceTest, DestroysAdaptationSession) {
 TEST_F(OnDeviceModelServiceTest, LoadsAdaptationWithPath) {
   FakeFile weights1("Adapt1");
   FakeFile weights2("Adapt2");
-  auto model = LoadModel(mojom::ModelBackendType::kApu);
+  auto model = LoadModel(ml::ModelBackendType::kApuBackend);
   auto adaptation1 = LoadAdaptation(*model, weights1.Path());
   EXPECT_THAT(GetResponses(*model, "foo"), ElementsAre("Input: foo\n"));
   EXPECT_THAT(GetResponses(*adaptation1, "foo"),
@@ -633,5 +636,12 @@ TEST_F(OnDeviceModelServiceTest, ClassifyTextSafety) {
   EXPECT_THAT(resp2->class_scores, ElementsAre(0.2, 0.2));
 }
 
+TEST_F(OnDeviceModelServiceTest, PerformanceHint) {
+  auto model = LoadModel(ml::ModelBackendType::kGpuBackend,
+                         ml::ModelPerformanceHint::kFastestInference);
+  EXPECT_THAT(GetResponses(*model, "foo"),
+              ElementsAre("Fastest inference\n", "Input: foo\n"));
+}
+
 }  // namespace
 }  // namespace on_device_model

services/on_device_model/public/cpp/test_support/fake_service.cc

@@ -130,6 +130,12 @@ void FakeOnDeviceSession::ExecuteImpl(
     mojom::InputOptionsPtr input,
     mojo::PendingRemote<mojom::StreamingResponder> response) {
   mojo::Remote<mojom::StreamingResponder> remote(std::move(response));
+  if (model_->performance_hint() ==
+      ml::ModelPerformanceHint::kFastestInference) {
+    auto chunk = mojom::ResponseChunk::New();
+    chunk->text = "Fastest inference\n";
+    remote->OnResponse(std::move(chunk));
+  }
   if (model_->data().base_weight != "0") {
     auto chunk = mojom::ResponseChunk::New();
     chunk->text = "Base model: " + model_->data().base_weight + "\n";
@@ -184,8 +190,11 @@ void FakeOnDeviceSession::AddContextInternal(
 }
 
 FakeOnDeviceModel::FakeOnDeviceModel(FakeOnDeviceServiceSettings* settings,
-                                     FakeOnDeviceModel::Data&& data)
-    : settings_(settings), data_(std::move(data)) {}
+                                     FakeOnDeviceModel::Data&& data,
+                                     ml::ModelPerformanceHint performance_hint)
+    : settings_(settings),
+      data_(std::move(data)),
+      performance_hint_(performance_hint) {}
 
 FakeOnDeviceModel::~FakeOnDeviceModel() = default;
 
@@ -221,8 +230,8 @@ void FakeOnDeviceModel::LoadAdaptation(
     LoadAdaptationCallback callback) {
   Data data = data_;
   data.adaptation_model_weight = ReadFile(params->assets.weights);
-  auto test_model =
-      std::make_unique<FakeOnDeviceModel>(settings_, std::move(data));
+  auto test_model = std::make_unique<FakeOnDeviceModel>(
+      settings_, std::move(data), ml::ModelPerformanceHint::kHighestQuality);
   model_adaptation_receivers_.Add(std::move(test_model), std::move(model));
   std::move(callback).Run(mojom::LoadModelResult::kSuccess);
 }
@@ -292,8 +301,8 @@ void FakeOnDeviceModelService::LoadModel(
   }
   FakeOnDeviceModel::Data data;
   data.base_weight = ReadFile(params->assets.weights);
-  auto test_model =
-      std::make_unique<FakeOnDeviceModel>(settings_, std::move(data));
+  auto test_model = std::make_unique<FakeOnDeviceModel>(
+      settings_, std::move(data), params->performance_hint);
   model_receivers_.Add(std::move(test_model), std::move(model));
   std::move(callback).Run(mojom::LoadModelResult::kSuccess);
 }

services/on_device_model/public/cpp/test_support/fake_service.h

@@ -112,7 +112,8 @@ class FakeOnDeviceModel : public mojom::OnDeviceModel {
     std::string adaptation_model_weight = "";
   };
   explicit FakeOnDeviceModel(FakeOnDeviceServiceSettings* settings,
-                             Data&& data);
+                             Data&& data,
+                             ml::ModelPerformanceHint performance_hint);
   ~FakeOnDeviceModel() override;
 
   // mojom::OnDeviceModel:
@@ -134,9 +135,14 @@ class FakeOnDeviceModel : public mojom::OnDeviceModel {
 
   const Data& data() const { return data_; }
 
+  ml::ModelPerformanceHint performance_hint() const {
+    return performance_hint_;
+  }
+
  private:
   raw_ptr<FakeOnDeviceServiceSettings> settings_;
   Data data_;
+  ml::ModelPerformanceHint performance_hint_;
 
   mojo::UniqueReceiverSet<mojom::Session> receivers_;
   mojo::UniqueReceiverSet<mojom::OnDeviceModel> model_adaptation_receivers_;

services/on_device_model/public/mojom/BUILD.gn

@@ -37,6 +37,14 @@ mojom("mojom") {
           mojom = "on_device_model.mojom.InputPiece"
           cpp = "::ml::InputPiece"
         },
+        {
+          mojom = "on_device_model.mojom.ModelBackendType"
+          cpp = "::ml::ModelBackendType"
+        },
+        {
+          mojom = "on_device_model.mojom.ModelPerformanceHint"
+          cpp = "::ml::ModelPerformanceHint"
+        },
       ]
       traits_headers = [
         "//services/on_device_model/public/cpp/adaptation_assets_mojom_traits.h",

services/on_device_model/public/mojom/on_device_model_service.mojom

@@ -36,6 +36,12 @@ enum ModelBackendType {
   kApu,
 };
 
+// Options for specifying the performance characteristics of the model to load.
+enum ModelPerformanceHint {
+  kHighestQuality,
+  kFastestInference,
+};
+
 // Params to describe the model to load.
 struct LoadModelParams {
   // Backend type of the model.
@@ -50,6 +56,10 @@ struct LoadModelParams {
 
   // List of adaptation ranks the model should support.
   array<uint32> adaptation_ranks;
+
+  // Chooses the performance characteristics of the model loaded. If only a
+  // single model is available, this field will do nothing.
+  ModelPerformanceHint performance_hint = kHighestQuality;
 };
 
 struct TextSafetyModelAssets {

tools/metrics/histograms/enums.xml

@@ -15298,6 +15298,7 @@ from previous Chrome versions.
   <int value="-2036127998" label="LocalWebApprovals:disabled"/>
   <int value="-2035845836" label="ArcExtendServiceAnrTimeout:enabled"/>
   <int value="-2035126988" label="enabled-new-style-notification"/>
+  <int value="-2034315473" label="OnDeviceModelPerformanceParams:disabled"/>
   <int value="-2034064186" label="EnableKeyboardBacklightToggle:disabled"/>
   <int value="-2033950090" label="AutofillNoLocalSaveOnUploadSuccess:disabled"/>
   <int value="-2033908928" label="NightLight:enabled"/>
@@ -22130,6 +22131,7 @@ from previous Chrome versions.
   <int value="705946076"
       label="ContextMenuPerformanceInfoAndRemoteHintFetching:disabled"/>
   <int value="706280254" label="StoragePressureEvent:enabled"/>
+  <int value="707243920" label="OnDeviceModelPerformanceParams:enabled"/>
   <int value="707463326" label="DynamicSafeAreaInsets:enabled"/>
   <int value="708015891"
       label="AutofillUpdateChromeSettingsLinkToGPayWeb:enabled"/>