diff --git a/camera/provider/2.4/vts/functional/VtsHalCameraProviderV2_4TargetTest.cpp b/camera/provider/2.4/vts/functional/VtsHalCameraProviderV2_4TargetTest.cpp
index c89d983..37d2973 100644
--- a/camera/provider/2.4/vts/functional/VtsHalCameraProviderV2_4TargetTest.cpp
+++ b/camera/provider/2.4/vts/functional/VtsHalCameraProviderV2_4TargetTest.cpp
@@ -935,7 +935,6 @@
             camera_metadata_ro_entry* streamConfigs,
             camera_metadata_ro_entry* maxResolutionStreamConfigs,
             const camera_metadata_t* staticMetadata);
-    static bool isColorCamera(const camera_metadata_t *metadata);
 
     static V3_2::DataspaceFlags getDataspace(PixelFormat format);
 
@@ -6187,141 +6186,6 @@
     }
 }
 
-// Test the multi-camera API requirement for Google Requirement Freeze S
-// Note that this requirement can only be partially tested. If a vendor
-// device doesn't expose a physical camera in any shape or form, there is no way
-// the test can catch it.
-TEST_P(CameraHidlTest, grfSMultiCameraTest) {
-    const int socGrfApi = property_get_int32("ro.board.first_api_level", /*default*/ -1);
-    if (socGrfApi < 31 /*S*/) {
-        // Non-GRF devices, or version < 31 Skip
-        ALOGI("%s: socGrfApi level is %d. Skipping", __FUNCTION__, socGrfApi);
-        return;
-    }
-
-    // Test that if more than one rear-facing color camera is
-    // supported, there must be at least one rear-facing logical camera.
-    hidl_vec<hidl_string> cameraDeviceNames = getCameraDeviceNames(mProvider);
-    // Back facing non-logical color cameras
-    std::set<std::string> rearColorCameras;
-    // Back facing logical cameras' physical camera Id sets
-    std::set<std::set<std::string>> rearPhysicalIds;
-    for (const auto& name : cameraDeviceNames) {
-        std::string cameraId;
-        int deviceVersion = getCameraDeviceVersionAndId(name, mProviderType, &cameraId);
-        switch (deviceVersion) {
-            case CAMERA_DEVICE_API_VERSION_3_7:
-            case CAMERA_DEVICE_API_VERSION_3_6:
-            case CAMERA_DEVICE_API_VERSION_3_5:
-            case CAMERA_DEVICE_API_VERSION_3_4:
-            case CAMERA_DEVICE_API_VERSION_3_3:
-            case CAMERA_DEVICE_API_VERSION_3_2: {
-                ::android::sp<::android::hardware::camera::device::V3_2::ICameraDevice> device3_x;
-                ALOGI("getCameraCharacteristics: Testing camera device %s", name.c_str());
-                Return<void> ret;
-                ret = mProvider->getCameraDeviceInterface_V3_x(
-                        name, [&](auto status, const auto& device) {
-                            ALOGI("getCameraDeviceInterface_V3_x returns status:%d", (int)status);
-                            ASSERT_EQ(Status::OK, status);
-                            ASSERT_NE(device, nullptr);
-                            device3_x = device;
-                        });
-                ASSERT_TRUE(ret.isOk());
-
-                ret = device3_x->getCameraCharacteristics([&](auto status, const auto& chars) {
-                    ASSERT_EQ(Status::OK, status);
-                    const camera_metadata_t* metadata = (camera_metadata_t*)chars.data();
-
-                    // Skip if this is not a color camera.
-                    if (!CameraHidlTest::isColorCamera(metadata)) {
-                        return;
-                    }
-
-                    // Check camera facing. Skip if facing is not BACK.
-                    // If this is not a logical camera, only note down
-                    // the camera ID, and skip.
-                    camera_metadata_ro_entry entry;
-                    int retcode = find_camera_metadata_ro_entry(
-                            metadata, ANDROID_LENS_FACING, &entry);
-                    ASSERT_EQ(retcode, 0);
-                    ASSERT_GT(entry.count, 0);
-                    uint8_t facing = entry.data.u8[0];
-                    bool isLogicalCamera = (isLogicalMultiCamera(metadata) == Status::OK);
-                    if (facing != ANDROID_LENS_FACING_BACK) {
-                        // Not BACK facing. Skip.
-                        return;
-                    }
-                    if (!isLogicalCamera) {
-                        rearColorCameras.insert(cameraId);
-                        return;
-                    }
-
-                    // Check logical camera's physical camera IDs for color
-                    // cameras.
-                    std::unordered_set<std::string> physicalCameraIds;
-                    Status s = getPhysicalCameraIds(metadata, &physicalCameraIds);
-                    ASSERT_EQ(Status::OK, s);
-                    rearPhysicalIds.emplace(physicalCameraIds.begin(), physicalCameraIds.end());
-                    for (const auto& physicalId : physicalCameraIds) {
-                        // Skip if the physicalId is publicly available
-                        for (auto& deviceName : cameraDeviceNames) {
-                            std::string publicVersion, publicId;
-                            ASSERT_TRUE(::matchDeviceName(deviceName, mProviderType,
-                                                          &publicVersion, &publicId));
-                            if (physicalId == publicId) {
-                                // Skip because public Ids will be iterated in outer loop.
-                                return;
-                            }
-                        }
-
-                        auto castResult = device::V3_5::ICameraDevice::castFrom(device3_x);
-                        ASSERT_TRUE(castResult.isOk());
-                        ::android::sp<::android::hardware::camera::device::V3_5::ICameraDevice>
-                                device3_5 = castResult;
-                        ASSERT_NE(device3_5, nullptr);
-
-                        // Check camera characteristics for hidden camera id
-                        Return<void> ret = device3_5->getPhysicalCameraCharacteristics(
-                                physicalId, [&](auto status, const auto& chars) {
-                            ASSERT_EQ(Status::OK, status);
-                            const camera_metadata_t* physicalMetadata =
-                                    (camera_metadata_t*)chars.data();
-
-                            if (CameraHidlTest::isColorCamera(physicalMetadata)) {
-                                rearColorCameras.insert(physicalId);
-                            }
-                        });
-                        ASSERT_TRUE(ret.isOk());
-                    }
-                });
-                ASSERT_TRUE(ret.isOk());
-            } break;
-            case CAMERA_DEVICE_API_VERSION_1_0: {
-                // Not applicable
-            } break;
-            default: {
-                ALOGE("%s: Unsupported device version %d", __func__, deviceVersion);
-                ADD_FAILURE();
-            } break;
-        }
-    }
-
-    // If there are more than one rear-facing color camera, a logical
-    // multi-camera must be defined consisting of all rear-facing color
-    // cameras.
-    if (rearColorCameras.size() > 1) {
-        bool hasRearLogical = false;
-        for (const auto& physicalIds : rearPhysicalIds) {
-            if (std::includes(physicalIds.begin(), physicalIds.end(),
-                    rearColorCameras.begin(), rearColorCameras.end())) {
-                hasRearLogical = true;
-                break;
-            }
-        }
-        ASSERT_TRUE(hasRearLogical);
-    }
-}
-
 // Retrieve all valid output stream resolutions from the camera
 // static characteristics.
 Status CameraHidlTest::getAvailableOutputStreams(const camera_metadata_t* staticMeta,
@@ -6794,23 +6658,6 @@
     return ret;
 }
 
-bool CameraHidlTest::isColorCamera(const camera_metadata_t *metadata) {
-    camera_metadata_ro_entry entry;
-    int retcode = find_camera_metadata_ro_entry(
-            metadata, ANDROID_REQUEST_AVAILABLE_CAPABILITIES, &entry);
-    if ((0 == retcode) && (entry.count > 0)) {
-        bool isBackwardCompatible = (std::find(entry.data.u8, entry.data.u8 + entry.count,
-                ANDROID_REQUEST_AVAILABLE_CAPABILITIES_BACKWARD_COMPATIBLE) !=
-                entry.data.u8 + entry.count);
-        bool isMonochrome = (std::find(entry.data.u8, entry.data.u8 + entry.count,
-                ANDROID_REQUEST_AVAILABLE_CAPABILITIES_MONOCHROME) !=
-                entry.data.u8 + entry.count);
-        bool isColor = isBackwardCompatible && !isMonochrome;
-        return isColor;
-    }
-    return false;
-}
-
 // Retrieve the reprocess input-output format map from the static
 // camera characteristics.
 Status CameraHidlTest::getZSLInputOutputMap(camera_metadata_t *staticMeta,
diff --git a/neuralnetworks/1.0/utils/Android.bp b/neuralnetworks/1.0/utils/Android.bp
index 31cdded..ad30e30 100644
--- a/neuralnetworks/1.0/utils/Android.bp
+++ b/neuralnetworks/1.0/utils/Android.bp
@@ -31,16 +31,11 @@
     export_include_dirs: ["include"],
     cflags: ["-Wthread-safety"],
     static_libs: [
+        "android.hardware.neuralnetworks@1.0",
         "libarect",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
     ],
-    shared_libs: [
-        "android.hardware.neuralnetworks@1.0",
-    ],
-    export_static_lib_headers: [
-        "neuralnetworks_utils_hal_common",
-    ],
     target: {
         android: {
             shared_libs: ["libnativewindow"],
@@ -55,19 +50,14 @@
     static_libs: [
         "android.hardware.neuralnetworks@1.0",
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
         "neuralnetworks_utils_hal_1_0",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
-        "android.hidl.memory@1.0",
         "libbase",
         "libcutils",
-        "libfmq",
         "libhidlbase",
-        "libhidlmemory",
         "liblog",
         "libutils",
     ],
diff --git a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Burst.h b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Burst.h
index 8bd2fbe..cef76c6 100644
--- a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Burst.h
+++ b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Burst.h
@@ -45,12 +45,15 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
   private:
     const nn::SharedPreparedModel kPreparedModel;
diff --git a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Device.h b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Device.h
index 0a6ca3e..d7c43ef 100644
--- a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Device.h
+++ b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/Device.h
@@ -65,8 +65,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
index bdb5b54..337c132 100644
--- a/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
+++ b/neuralnetworks/1.0/utils/include/nnapi/hal/1.0/PreparedModel.h
@@ -49,18 +49,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
diff --git a/neuralnetworks/1.0/utils/src/Burst.cpp b/neuralnetworks/1.0/utils/src/Burst.cpp
index 1284721..3642bc6 100644
--- a/neuralnetworks/1.0/utils/src/Burst.cpp
+++ b/neuralnetworks/1.0/utils/src/Burst.cpp
@@ -50,15 +50,20 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
-    return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
+    return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration, hints,
+                                   extensionNameToPrefix);
 }
 
 nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
-    return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration);
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
+    return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration, hints,
+                                                   extensionNameToPrefix);
 }
 
 }  // namespace android::hardware::neuralnetworks::V1_0::utils
diff --git a/neuralnetworks/1.0/utils/src/Device.cpp b/neuralnetworks/1.0/utils/src/Device.cpp
index b0c236e..620d040 100644
--- a/neuralnetworks/1.0/utils/src/Device.cpp
+++ b/neuralnetworks/1.0/utils/src/Device.cpp
@@ -143,7 +143,9 @@
 nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
         const nn::Model& model, nn::ExecutionPreference /*preference*/, nn::Priority /*priority*/,
         nn::OptionalTimePoint /*deadline*/, const std::vector<nn::SharedHandle>& /*modelCache*/,
-        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/) const {
+        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that model is ready for IPC.
     std::optional<nn::Model> maybeModelInShared;
     const nn::Model& modelInShared =
diff --git a/neuralnetworks/1.0/utils/src/PreparedModel.cpp b/neuralnetworks/1.0/utils/src/PreparedModel.cpp
index 00e7d22..b8055fc 100644
--- a/neuralnetworks/1.0/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.0/utils/src/PreparedModel.cpp
@@ -59,7 +59,9 @@
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
         const nn::Request& request, nn::MeasureTiming /*measure*/,
         const nn::OptionalTimePoint& /*deadline*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -94,19 +96,22 @@
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFenced(const nn::Request& /*request*/,
-                             const std::vector<nn::SyncFence>& /*waitFor*/,
-                             nn::MeasureTiming /*measure*/,
-                             const nn::OptionalTimePoint& /*deadline*/,
-                             const nn::OptionalDuration& /*loopTimeoutDuration*/,
-                             const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+PreparedModel::executeFenced(
+        const nn::Request& /*request*/, const std::vector<nn::SyncFence>& /*waitFor*/,
+        nn::MeasureTiming /*measure*/, const nn::OptionalTimePoint& /*deadline*/,
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const nn::OptionalDuration& /*timeoutDurationAfterFence*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
            << "IPreparedModel::executeFenced is not supported on 1.0 HAL service";
 }
 
 nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming /*measure*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
diff --git a/neuralnetworks/1.0/utils/test/DeviceTest.cpp b/neuralnetworks/1.0/utils/test/DeviceTest.cpp
index 83e555f..9e9db16 100644
--- a/neuralnetworks/1.0/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/1.0/utils/test/DeviceTest.cpp
@@ -380,7 +380,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -399,7 +399,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -417,7 +417,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -435,7 +435,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -452,7 +452,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -469,7 +469,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -488,7 +488,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp b/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
index 7820c06..e03a98d 100644
--- a/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/1.0/utils/test/PreparedModelTest.cpp
@@ -121,7 +121,7 @@
             .WillOnce(Invoke(makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::NONE)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -138,7 +138,7 @@
                                          V1_0::ErrorStatus::GENERAL_FAILURE)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -155,7 +155,7 @@
                     makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::GENERAL_FAILURE)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -171,7 +171,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -187,7 +187,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -205,7 +205,7 @@
     EXPECT_CALL(*mockPreparedModel, execute(_, _)).Times(1).WillOnce(InvokeWithoutArgs(ret));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -218,7 +218,7 @@
     const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -235,7 +235,7 @@
             .WillRepeatedly(Invoke(makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::NONE)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -258,7 +258,7 @@
                                          V1_0::ErrorStatus::GENERAL_FAILURE)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -279,7 +279,7 @@
                     makeExecute(V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::GENERAL_FAILURE)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -299,7 +299,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -319,7 +319,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -341,7 +341,7 @@
     EXPECT_CALL(*mockPreparedModel, execute(_, _)).Times(1).WillOnce(InvokeWithoutArgs(ret));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -358,7 +358,7 @@
     const auto preparedModel = PreparedModel::create(mockPreparedModel).value();
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
diff --git a/neuralnetworks/1.1/utils/Android.bp b/neuralnetworks/1.1/utils/Android.bp
index 737ff58..4b8999f 100644
--- a/neuralnetworks/1.1/utils/Android.bp
+++ b/neuralnetworks/1.1/utils/Android.bp
@@ -31,17 +31,12 @@
     export_include_dirs: ["include"],
     cflags: ["-Wthread-safety"],
     static_libs: [
+        "android.hardware.neuralnetworks@1.0",
+        "android.hardware.neuralnetworks@1.1",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
         "neuralnetworks_utils_hal_1_0",
     ],
-    shared_libs: [
-        "android.hardware.neuralnetworks@1.0",
-        "android.hardware.neuralnetworks@1.1",
-    ],
-    export_static_lib_headers: [
-        "neuralnetworks_utils_hal_common",
-    ],
 }
 
 cc_test {
@@ -52,20 +47,15 @@
         "android.hardware.neuralnetworks@1.0",
         "android.hardware.neuralnetworks@1.1",
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
         "neuralnetworks_utils_hal_1_0",
         "neuralnetworks_utils_hal_1_1",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
-        "android.hidl.memory@1.0",
         "libbase",
         "libcutils",
-        "libfmq",
         "libhidlbase",
-        "libhidlmemory",
         "liblog",
         "libutils",
     ],
diff --git a/neuralnetworks/1.1/utils/include/nnapi/hal/1.1/Device.h b/neuralnetworks/1.1/utils/include/nnapi/hal/1.1/Device.h
index d6bd36a..38ca138 100644
--- a/neuralnetworks/1.1/utils/include/nnapi/hal/1.1/Device.h
+++ b/neuralnetworks/1.1/utils/include/nnapi/hal/1.1/Device.h
@@ -64,8 +64,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/1.1/utils/src/Device.cpp b/neuralnetworks/1.1/utils/src/Device.cpp
index 3effa84..28f3276 100644
--- a/neuralnetworks/1.1/utils/src/Device.cpp
+++ b/neuralnetworks/1.1/utils/src/Device.cpp
@@ -143,7 +143,9 @@
 nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority /*priority*/,
         nn::OptionalTimePoint /*deadline*/, const std::vector<nn::SharedHandle>& /*modelCache*/,
-        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/) const {
+        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that model is ready for IPC.
     std::optional<nn::Model> maybeModelInShared;
     const nn::Model& modelInShared =
diff --git a/neuralnetworks/1.1/utils/test/DeviceTest.cpp b/neuralnetworks/1.1/utils/test/DeviceTest.cpp
index 2248da6..8ab87bc 100644
--- a/neuralnetworks/1.1/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/1.1/utils/test/DeviceTest.cpp
@@ -390,7 +390,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -409,7 +409,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -427,7 +427,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -445,7 +445,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -462,7 +462,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -479,7 +479,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -498,7 +498,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/1.2/utils/Android.bp b/neuralnetworks/1.2/utils/Android.bp
index 4eefb0f..4c5f065 100644
--- a/neuralnetworks/1.2/utils/Android.bp
+++ b/neuralnetworks/1.2/utils/Android.bp
@@ -31,19 +31,14 @@
     export_include_dirs: ["include"],
     cflags: ["-Wthread-safety"],
     static_libs: [
-        "neuralnetworks_types",
-        "neuralnetworks_utils_hal_common",
-        "neuralnetworks_utils_hal_1_0",
-        "neuralnetworks_utils_hal_1_1",
-    ],
-    shared_libs: [
         "android.hardware.neuralnetworks@1.0",
         "android.hardware.neuralnetworks@1.1",
         "android.hardware.neuralnetworks@1.2",
         "libfmq",
-    ],
-    export_static_lib_headers: [
+        "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
+        "neuralnetworks_utils_hal_1_0",
+        "neuralnetworks_utils_hal_1_1",
     ],
     product_variables: {
         debuggable: { // eng and userdebug builds
@@ -71,7 +66,6 @@
         "android.hardware.neuralnetworks@1.1",
         "android.hardware.neuralnetworks@1.2",
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
         "neuralnetworks_utils_hal_1_0",
@@ -79,13 +73,10 @@
         "neuralnetworks_utils_hal_1_2",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
-        "android.hidl.memory@1.0",
         "libbase",
         "libcutils",
         "libfmq",
         "libhidlbase",
-        "libhidlmemory",
         "liblog",
         "libutils",
     ],
diff --git a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Burst.h b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Burst.h
index ac9411c..1b28476 100644
--- a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Burst.h
+++ b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Burst.h
@@ -170,13 +170,16 @@
     // See IBurst::execute for information on this method.
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     // See IBurst::createReusableExecution for information on this method.
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     // If fallback is not nullptr, this method will invoke the fallback function to try another
     // execution path if the packet could not be sent. Otherwise, failing to send the packet will
diff --git a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Conversions.h b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Conversions.h
index c3348aa..4f13adc 100644
--- a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Conversions.h
+++ b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Conversions.h
@@ -37,7 +37,7 @@
 GeneralResult<Operand::ExtraParams> unvalidatedConvert(
         const hal::V1_2::Operand::ExtraParams& extraParams);
 GeneralResult<Model> unvalidatedConvert(const hal::V1_2::Model& model);
-GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
+GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
         const hal::V1_2::Model::ExtensionNameAndPrefix& extensionNameAndPrefix);
 GeneralResult<OutputShape> unvalidatedConvert(const hal::V1_2::OutputShape& outputShape);
 GeneralResult<MeasureTiming> unvalidatedConvert(const hal::V1_2::MeasureTiming& measureTiming);
@@ -78,7 +78,7 @@
         const nn::Operand::ExtraParams& extraParams);
 nn::GeneralResult<Model> unvalidatedConvert(const nn::Model& model);
 nn::GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
-        const nn::Model::ExtensionNameAndPrefix& extensionNameAndPrefix);
+        const nn::ExtensionNameAndPrefix& extensionNameAndPrefix);
 nn::GeneralResult<OutputShape> unvalidatedConvert(const nn::OutputShape& outputShape);
 nn::GeneralResult<MeasureTiming> unvalidatedConvert(const nn::MeasureTiming& measureTiming);
 nn::GeneralResult<Timing> unvalidatedConvert(const nn::Timing& timing);
diff --git a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Device.h b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Device.h
index e7ac172..d92cf50 100644
--- a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Device.h
+++ b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/Device.h
@@ -83,8 +83,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/PreparedModel.h b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/PreparedModel.h
index 1150e5e..72a5b2f 100644
--- a/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/PreparedModel.h
+++ b/neuralnetworks/1.2/utils/include/nnapi/hal/1.2/PreparedModel.h
@@ -49,18 +49,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
diff --git a/neuralnetworks/1.2/utils/src/Burst.cpp b/neuralnetworks/1.2/utils/src/Burst.cpp
index 911fbfa..23e8070 100644
--- a/neuralnetworks/1.2/utils/src/Burst.cpp
+++ b/neuralnetworks/1.2/utils/src/Burst.cpp
@@ -305,8 +305,9 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // This is the first point when we know an execution is occurring, so begin to collect
     // systraces. Note that the first point we can begin collecting systraces in
     // ExecutionBurstServer is when the RequestChannelReceiver realizes there is data in the FMQ, so
@@ -317,7 +318,7 @@
     // fall back to another execution path
     if (!compliantVersion(request).ok()) {
         // fallback to another execution path if the packet could not be sent
-        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration, {}, {});
     }
 
     // ensure that request is ready for IPC
@@ -346,7 +347,7 @@
     // send request packet
     const auto requestPacket = serialize(hidlRequest, hidlMeasure, slots);
     const auto fallback = [this, &request, measure, &deadline, &loopTimeoutDuration] {
-        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration);
+        return kPreparedModel->execute(request, measure, deadline, loopTimeoutDuration, {}, {});
     };
     return executeInternal(requestPacket, relocation, fallback);
 }
@@ -354,14 +355,17 @@
 // See IBurst::createReusableExecution for information on this method.
 nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     NNTRACE_RT(NNTRACE_PHASE_EXECUTION, "Burst::createReusableExecution");
 
     // if the request is valid but of a higher version than what's supported in burst execution,
     // fall back to another execution path
     if (!compliantVersion(request).ok()) {
         // fallback to another execution path if the packet could not be sent
-        return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration);
+        return kPreparedModel->createReusableExecution(request, measure, loopTimeoutDuration, {},
+                                                       {});
     }
 
     // ensure that request is ready for IPC
diff --git a/neuralnetworks/1.2/utils/src/Conversions.cpp b/neuralnetworks/1.2/utils/src/Conversions.cpp
index 838d9c4..62ec2ed 100644
--- a/neuralnetworks/1.2/utils/src/Conversions.cpp
+++ b/neuralnetworks/1.2/utils/src/Conversions.cpp
@@ -212,9 +212,9 @@
     };
 }
 
-GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
+GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
         const hal::V1_2::Model::ExtensionNameAndPrefix& extensionNameAndPrefix) {
-    return Model::ExtensionNameAndPrefix{
+    return ExtensionNameAndPrefix{
             .name = extensionNameAndPrefix.name,
             .prefix = extensionNameAndPrefix.prefix,
     };
@@ -495,7 +495,7 @@
 }
 
 nn::GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
-        const nn::Model::ExtensionNameAndPrefix& extensionNameAndPrefix) {
+        const nn::ExtensionNameAndPrefix& extensionNameAndPrefix) {
     return Model::ExtensionNameAndPrefix{
             .name = extensionNameAndPrefix.name,
             .prefix = extensionNameAndPrefix.prefix,
diff --git a/neuralnetworks/1.2/utils/src/Device.cpp b/neuralnetworks/1.2/utils/src/Device.cpp
index e7acecd..3a58d2c 100644
--- a/neuralnetworks/1.2/utils/src/Device.cpp
+++ b/neuralnetworks/1.2/utils/src/Device.cpp
@@ -236,7 +236,9 @@
 nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority /*priority*/,
         nn::OptionalTimePoint /*deadline*/, const std::vector<nn::SharedHandle>& modelCache,
-        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that model is ready for IPC.
     std::optional<nn::Model> maybeModelInShared;
     const nn::Model& modelInShared =
diff --git a/neuralnetworks/1.2/utils/src/PreparedModel.cpp b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
index 6df3df3..feb3951 100644
--- a/neuralnetworks/1.2/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.2/utils/src/PreparedModel.cpp
@@ -91,7 +91,9 @@
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
         const nn::Request& request, nn::MeasureTiming measure,
         const nn::OptionalTimePoint& /*deadline*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -123,19 +125,22 @@
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFenced(const nn::Request& /*request*/,
-                             const std::vector<nn::SyncFence>& /*waitFor*/,
-                             nn::MeasureTiming /*measure*/,
-                             const nn::OptionalTimePoint& /*deadline*/,
-                             const nn::OptionalDuration& /*loopTimeoutDuration*/,
-                             const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+PreparedModel::executeFenced(
+        const nn::Request& /*request*/, const std::vector<nn::SyncFence>& /*waitFor*/,
+        nn::MeasureTiming /*measure*/, const nn::OptionalTimePoint& /*deadline*/,
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const nn::OptionalDuration& /*timeoutDurationAfterFence*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
            << "IPreparedModel::executeFenced is not supported on 1.2 HAL service";
 }
 
 nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
diff --git a/neuralnetworks/1.2/utils/test/DeviceTest.cpp b/neuralnetworks/1.2/utils/test/DeviceTest.cpp
index 1dc6285..0d8c141 100644
--- a/neuralnetworks/1.2/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/1.2/utils/test/DeviceTest.cpp
@@ -636,7 +636,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -655,7 +655,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -673,7 +673,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -691,7 +691,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -708,7 +708,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -725,7 +725,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -746,7 +746,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/1.2/utils/test/PreparedModelTest.cpp b/neuralnetworks/1.2/utils/test/PreparedModelTest.cpp
index 5e2ad79..a5ec9d3 100644
--- a/neuralnetworks/1.2/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/1.2/utils/test/PreparedModelTest.cpp
@@ -154,7 +154,7 @@
             .WillOnce(Invoke(makeExecuteSynchronously(V1_0::ErrorStatus::NONE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -172,7 +172,7 @@
                     makeExecuteSynchronously(V1_0::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -189,7 +189,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -206,7 +206,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -224,7 +224,7 @@
                                                        V1_0::ErrorStatus::NONE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -243,7 +243,7 @@
                                                        kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -261,7 +261,7 @@
                     V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -278,7 +278,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -295,7 +295,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -314,7 +314,7 @@
     EXPECT_CALL(*mockPreparedModel, execute_1_2(_, _, _)).Times(1).WillOnce(InvokeWithoutArgs(ret));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -328,7 +328,7 @@
             PreparedModel::create(mockPreparedModel, /*executeSynchronously=*/true).value();
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -347,7 +347,7 @@
                     Invoke(makeExecuteSynchronously(V1_0::ErrorStatus::NONE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -371,7 +371,7 @@
                     makeExecuteSynchronously(V1_0::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -392,7 +392,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -413,7 +413,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -436,7 +436,7 @@
                     V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::NONE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -461,7 +461,7 @@
                                                        kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -483,7 +483,7 @@
                     V1_0::ErrorStatus::NONE, V1_0::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -504,7 +504,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -525,7 +525,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -548,7 +548,7 @@
     EXPECT_CALL(*mockPreparedModel, execute_1_2(_, _, _)).Times(1).WillOnce(InvokeWithoutArgs(ret));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -566,7 +566,7 @@
             PreparedModel::create(mockPreparedModel, /*executeSynchronously=*/true).value();
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
diff --git a/neuralnetworks/1.3/utils/Android.bp b/neuralnetworks/1.3/utils/Android.bp
index 7acb4fc..c512dda 100644
--- a/neuralnetworks/1.3/utils/Android.bp
+++ b/neuralnetworks/1.3/utils/Android.bp
@@ -31,21 +31,16 @@
     export_include_dirs: ["include"],
     cflags: ["-Wthread-safety"],
     static_libs: [
-        "neuralnetworks_types",
-        "neuralnetworks_utils_hal_common",
-        "neuralnetworks_utils_hal_1_0",
-        "neuralnetworks_utils_hal_1_1",
-        "neuralnetworks_utils_hal_1_2",
-    ],
-    shared_libs: [
         "android.hardware.neuralnetworks@1.0",
         "android.hardware.neuralnetworks@1.1",
         "android.hardware.neuralnetworks@1.2",
         "android.hardware.neuralnetworks@1.3",
         "libfmq",
-    ],
-    export_static_lib_headers: [
+        "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
+        "neuralnetworks_utils_hal_1_0",
+        "neuralnetworks_utils_hal_1_1",
+        "neuralnetworks_utils_hal_1_2",
     ],
     target: {
         host: {
@@ -69,7 +64,6 @@
         "android.hardware.neuralnetworks@1.2",
         "android.hardware.neuralnetworks@1.3",
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
         "neuralnetworks_utils_hal_1_0",
@@ -78,13 +72,10 @@
         "neuralnetworks_utils_hal_1_3",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
-        "android.hidl.memory@1.0",
         "libbase",
         "libcutils",
         "libfmq",
         "libhidlbase",
-        "libhidlmemory",
         "liblog",
         "libutils",
     ],
diff --git a/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/Device.h b/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/Device.h
index c3c6fc4..cf5e5ea0 100644
--- a/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/Device.h
+++ b/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/Device.h
@@ -66,8 +66,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/PreparedModel.h b/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/PreparedModel.h
index 480438d..124cc43 100644
--- a/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/PreparedModel.h
+++ b/neuralnetworks/1.3/utils/include/nnapi/hal/1.3/PreparedModel.h
@@ -48,18 +48,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
diff --git a/neuralnetworks/1.3/utils/src/Conversions.cpp b/neuralnetworks/1.3/utils/src/Conversions.cpp
index a1d414c..09e9d80 100644
--- a/neuralnetworks/1.3/utils/src/Conversions.cpp
+++ b/neuralnetworks/1.3/utils/src/Conversions.cpp
@@ -396,7 +396,7 @@
 }
 
 nn::GeneralResult<V1_2::Model::ExtensionNameAndPrefix> unvalidatedConvert(
-        const nn::Model::ExtensionNameAndPrefix& extensionNameAndPrefix) {
+        const nn::ExtensionNameAndPrefix& extensionNameAndPrefix) {
     return V1_2::utils::unvalidatedConvert(extensionNameAndPrefix);
 }
 
diff --git a/neuralnetworks/1.3/utils/src/Device.cpp b/neuralnetworks/1.3/utils/src/Device.cpp
index 9517fda..824cec6 100644
--- a/neuralnetworks/1.3/utils/src/Device.cpp
+++ b/neuralnetworks/1.3/utils/src/Device.cpp
@@ -187,7 +187,9 @@
 nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
         nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that model is ready for IPC.
     std::optional<nn::Model> maybeModelInShared;
     const nn::Model& modelInShared =
diff --git a/neuralnetworks/1.3/utils/src/PreparedModel.cpp b/neuralnetworks/1.3/utils/src/PreparedModel.cpp
index ce977e5..b92f877 100644
--- a/neuralnetworks/1.3/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/1.3/utils/src/PreparedModel.cpp
@@ -135,8 +135,9 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -174,10 +175,13 @@
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFenced(const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
-                             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
-                             const nn::OptionalDuration& loopTimeoutDuration,
-                             const nn::OptionalDuration& timeoutDurationAfterFence) const {
+PreparedModel::executeFenced(
+        const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
+        nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const nn::OptionalDuration& timeoutDurationAfterFence,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -230,7 +234,9 @@
 
 nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
diff --git a/neuralnetworks/1.3/utils/test/DeviceTest.cpp b/neuralnetworks/1.3/utils/test/DeviceTest.cpp
index 7eba4bc..6f48837 100644
--- a/neuralnetworks/1.3/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/1.3/utils/test/DeviceTest.cpp
@@ -658,7 +658,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -677,7 +677,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -695,7 +695,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -713,7 +713,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -730,7 +730,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -747,7 +747,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -768,7 +768,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/1.3/utils/test/PreparedModelTest.cpp b/neuralnetworks/1.3/utils/test/PreparedModelTest.cpp
index 6dbbd6b..51b5d29 100644
--- a/neuralnetworks/1.3/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/1.3/utils/test/PreparedModelTest.cpp
@@ -182,7 +182,7 @@
             .WillOnce(Invoke(makeExecuteSynchronously(V1_3::ErrorStatus::NONE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -200,7 +200,7 @@
                     makeExecuteSynchronously(V1_3::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -217,7 +217,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -234,7 +234,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -252,7 +252,7 @@
                                                        V1_3::ErrorStatus::NONE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -271,7 +271,7 @@
                                                        kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -289,7 +289,7 @@
                     V1_3::ErrorStatus::NONE, V1_3::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -306,7 +306,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -323,7 +323,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -344,7 +344,7 @@
             .WillOnce(InvokeWithoutArgs(ret));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -366,7 +366,7 @@
             .WillOnce(Invoke(makeExecuteFencedReturn(V1_3::ErrorStatus::NONE, {}, mockCallback)));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -396,7 +396,7 @@
             .WillOnce(Invoke(makeExecuteFencedReturn(V1_3::ErrorStatus::NONE, {}, mockCallback)));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -422,7 +422,7 @@
                     makeExecuteFencedReturn(V1_3::ErrorStatus::GENERAL_FAILURE, {}, nullptr)));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -439,7 +439,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -456,7 +456,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -475,7 +475,7 @@
                     Invoke(makeExecuteSynchronously(V1_3::ErrorStatus::NONE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -499,7 +499,7 @@
                     makeExecuteSynchronously(V1_3::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -520,7 +520,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -541,7 +541,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -564,7 +564,7 @@
                     V1_3::ErrorStatus::NONE, V1_3::ErrorStatus::NONE, {}, kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -589,7 +589,7 @@
                                                        kNoTiming)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -611,7 +611,7 @@
                     V1_3::ErrorStatus::NONE, V1_3::ErrorStatus::GENERAL_FAILURE, {}, kNoTiming)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -628,7 +628,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -649,7 +649,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -674,7 +674,7 @@
             .WillOnce(InvokeWithoutArgs(ret));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -702,7 +702,7 @@
                     Invoke(makeExecuteFencedReturn(V1_3::ErrorStatus::NONE, {}, mockCallback)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -738,7 +738,7 @@
             .WillOnce(Invoke(makeExecuteFencedReturn(V1_3::ErrorStatus::NONE, {}, mockCallback)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -768,7 +768,7 @@
                     makeExecuteFencedReturn(V1_3::ErrorStatus::GENERAL_FAILURE, {}, nullptr)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -789,7 +789,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -810,7 +810,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionConfig.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionConfig.aidl
new file mode 100644
index 0000000..cb85743
--- /dev/null
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionConfig.aidl
@@ -0,0 +1,41 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+///////////////////////////////////////////////////////////////////////////////
+// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE.                          //
+///////////////////////////////////////////////////////////////////////////////
+
+// This file is a snapshot of an AIDL file. Do not edit it manually. There are
+// two cases:
+// 1). this is a frozen version file - do not edit this in any case.
+// 2). this is a 'current' file. If you make a backwards compatible change to
+//     the interface (from the latest frozen version), the build system will
+//     prompt you to update this file with `m <name>-update-api`.
+//
+// You must not make a backward incompatible change to any AIDL file built
+// with the aidl_interface module type with versions property set. The module
+// type is used to build AIDL files in a way that they can be used across
+// independently updatable components of the system. If a device is shipped
+// with such a backward incompatible change, it has a high risk of breaking
+// later when a module using the interface is updated, e.g., Mainline modules.
+
+package android.hardware.neuralnetworks;
+@VintfStability
+parcelable ExecutionConfig {
+  boolean measureTiming;
+  long loopTimeoutDurationNs;
+  android.hardware.neuralnetworks.TokenValuePair[] executionHints;
+  android.hardware.neuralnetworks.ExtensionNameAndPrefix[] extensionNameToPrefix;
+}
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBurst.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBurst.aidl
index eb3d0b0..461fdfa 100644
--- a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBurst.aidl
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBurst.aidl
@@ -36,4 +36,5 @@
 interface IBurst {
   android.hardware.neuralnetworks.ExecutionResult executeSynchronously(in android.hardware.neuralnetworks.Request request, in long[] memoryIdentifierTokens, in boolean measureTiming, in long deadlineNs, in long loopTimeoutDurationNs);
   void releaseMemoryResource(in long memoryIdentifierToken);
+  android.hardware.neuralnetworks.ExecutionResult executeSynchronouslyWithConfig(in android.hardware.neuralnetworks.Request request, in long[] memoryIdentifierTokens, in android.hardware.neuralnetworks.ExecutionConfig config, in long deadlineNs);
 }
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl
index c9c67f2..c0fba47 100644
--- a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl
@@ -43,6 +43,7 @@
   String getVersionString();
   void prepareModel(in android.hardware.neuralnetworks.Model model, in android.hardware.neuralnetworks.ExecutionPreference preference, in android.hardware.neuralnetworks.Priority priority, in long deadlineNs, in ParcelFileDescriptor[] modelCache, in ParcelFileDescriptor[] dataCache, in byte[] token, in android.hardware.neuralnetworks.IPreparedModelCallback callback);
   void prepareModelFromCache(in long deadlineNs, in ParcelFileDescriptor[] modelCache, in ParcelFileDescriptor[] dataCache, in byte[] token, in android.hardware.neuralnetworks.IPreparedModelCallback callback);
+  void prepareModelWithConfig(in android.hardware.neuralnetworks.Model model, in android.hardware.neuralnetworks.PrepareModelConfig config, in android.hardware.neuralnetworks.IPreparedModelCallback callback);
   const int BYTE_SIZE_OF_CACHE_TOKEN = 32;
   const int MAX_NUMBER_OF_CACHE_FILES = 32;
   const int EXTENSION_TYPE_HIGH_BITS_PREFIX = 15;
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl
index f899567..fb0c372 100644
--- a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl
@@ -37,7 +37,9 @@
   android.hardware.neuralnetworks.ExecutionResult executeSynchronously(in android.hardware.neuralnetworks.Request request, in boolean measureTiming, in long deadlineNs, in long loopTimeoutDurationNs);
   android.hardware.neuralnetworks.FencedExecutionResult executeFenced(in android.hardware.neuralnetworks.Request request, in ParcelFileDescriptor[] waitFor, in boolean measureTiming, in long deadlineNs, in long loopTimeoutDurationNs, in long durationNs);
   android.hardware.neuralnetworks.IBurst configureExecutionBurst();
-  android.hardware.neuralnetworks.IExecution createReusableExecution(in android.hardware.neuralnetworks.Request request, in boolean measureTiming, in long loopTimeoutDurationNs);
+  android.hardware.neuralnetworks.IExecution createReusableExecution(in android.hardware.neuralnetworks.Request request, in android.hardware.neuralnetworks.ExecutionConfig config);
+  android.hardware.neuralnetworks.ExecutionResult executeSynchronouslyWithConfig(in android.hardware.neuralnetworks.Request request, in android.hardware.neuralnetworks.ExecutionConfig config, in long deadlineNs);
+  android.hardware.neuralnetworks.FencedExecutionResult executeFencedWithConfig(in android.hardware.neuralnetworks.Request request, in ParcelFileDescriptor[] waitFor, in android.hardware.neuralnetworks.ExecutionConfig config, in long deadlineNs, in long durationNs);
   const long DEFAULT_LOOP_TIMEOUT_DURATION_NS = 2000000000;
   const long MAXIMUM_LOOP_TIMEOUT_DURATION_NS = 15000000000;
 }
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl
new file mode 100644
index 0000000..85c924f
--- /dev/null
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PrepareModelConfig.aidl
@@ -0,0 +1,45 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+///////////////////////////////////////////////////////////////////////////////
+// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE.                          //
+///////////////////////////////////////////////////////////////////////////////
+
+// This file is a snapshot of an AIDL file. Do not edit it manually. There are
+// two cases:
+// 1). this is a frozen version file - do not edit this in any case.
+// 2). this is a 'current' file. If you make a backwards compatible change to
+//     the interface (from the latest frozen version), the build system will
+//     prompt you to update this file with `m <name>-update-api`.
+//
+// You must not make a backward incompatible change to any AIDL file built
+// with the aidl_interface module type with versions property set. The module
+// type is used to build AIDL files in a way that they can be used across
+// independently updatable components of the system. If a device is shipped
+// with such a backward incompatible change, it has a high risk of breaking
+// later when a module using the interface is updated, e.g., Mainline modules.
+
+package android.hardware.neuralnetworks;
+@VintfStability
+parcelable PrepareModelConfig {
+  android.hardware.neuralnetworks.ExecutionPreference preference;
+  android.hardware.neuralnetworks.Priority priority;
+  long deadlineNs;
+  ParcelFileDescriptor[] modelCache;
+  ParcelFileDescriptor[] dataCache;
+  byte[] cacheToken;
+  android.hardware.neuralnetworks.TokenValuePair[] compilationHints;
+  android.hardware.neuralnetworks.ExtensionNameAndPrefix[] extensionNameToPrefix;
+}
diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/TokenValuePair.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/TokenValuePair.aidl
new file mode 100644
index 0000000..e477d6e
--- /dev/null
+++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/TokenValuePair.aidl
@@ -0,0 +1,39 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+///////////////////////////////////////////////////////////////////////////////
+// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE.                          //
+///////////////////////////////////////////////////////////////////////////////
+
+// This file is a snapshot of an AIDL file. Do not edit it manually. There are
+// two cases:
+// 1). this is a frozen version file - do not edit this in any case.
+// 2). this is a 'current' file. If you make a backwards compatible change to
+//     the interface (from the latest frozen version), the build system will
+//     prompt you to update this file with `m <name>-update-api`.
+//
+// You must not make a backward incompatible change to any AIDL file built
+// with the aidl_interface module type with versions property set. The module
+// type is used to build AIDL files in a way that they can be used across
+// independently updatable components of the system. If a device is shipped
+// with such a backward incompatible change, it has a high risk of breaking
+// later when a module using the interface is updated, e.g., Mainline modules.
+
+package android.hardware.neuralnetworks;
+@VintfStability
+parcelable TokenValuePair {
+  int token;
+  byte[] value;
+}
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionConfig.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionConfig.aidl
new file mode 100644
index 0000000..00f1e11
--- /dev/null
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionConfig.aidl
@@ -0,0 +1,60 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package android.hardware.neuralnetworks;
+
+import android.hardware.neuralnetworks.ExtensionNameAndPrefix;
+import android.hardware.neuralnetworks.TokenValuePair;
+
+/**
+ * A type that is used to represent all configuration related to
+ * an Execution.
+ */
+@VintfStability
+parcelable ExecutionConfig {
+    /**
+     * Specifies whether or not to measure duration of the execution.
+     * For {@link IPreparedModel::executeSynchronouslyWithConfig}, the duration runs from the time
+     * the driver sees the corresponding call to the execute function to the time the driver returns
+     * from the function. For {@link IPreparedModel::executeFencedWithConfig}, please refer to
+     * {@link IPreparedModelCallback} for details.
+     */
+    boolean measureTiming;
+    /**
+     * The maximum amount of time in nanoseconds that should be spent
+     * executing a {@link OperationType::WHILE} operation. If a loop
+     * condition model does not output false within this duration,
+     * the execution must be aborted. If -1 is provided, the maximum
+     * amount of time is {@link DEFAULT_LOOP_TIMEOUT_DURATION_NS}.
+     * Other negative values are invalid. When provided, the duration
+     * must not exceed {@link MAXIMUM_LOOP_TIMEOUT_DURATION_NS}.
+     */
+    long loopTimeoutDurationNs;
+    /**
+     * A vector of token / value pairs represent vendor specific
+     * execution hints or metadata. The provided TokenValuePairs must not
+     * contain the same token twice. The driver must validate the
+     * data and ignore invalid hints. It is up to the driver to
+     * decide whether to respect the provided hints or not.
+     */
+    TokenValuePair[] executionHints;
+    /**
+     * The mapping between extension names and prefixes of token values.
+     * The driver must ignore the corresponding execution hint, if
+     * the extension is not supported.
+     */
+    ExtensionNameAndPrefix[] extensionNameToPrefix;
+}
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl
index 20109bd..9f70a53 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl
@@ -20,6 +20,10 @@
 
 /**
  * Information about an extension.
+ *
+ * The extension can provide zero or more operation types (which are not enumerated), zero or more
+ * operand types (which are enumerated in {@link Extension::operandTypes}, and compilation and
+ * execution hints (which are not enumerated).
  */
 @VintfStability
 parcelable Extension {
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl
index 29be93f..6c296e0 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl
@@ -17,7 +17,8 @@
 package android.hardware.neuralnetworks;
 
 /**
- * The mapping between extension names and prefixes of operand and operation type values.
+ * The mapping between extension names and prefixes of values like operand and operation type, and
+ * token in {@link TokenValuePair}.
  *
  * An operand or operation whose numeric type value is above {@link IDevice::OPERAND_TYPE_BASE_MAX}
  * or {@link IDevice::OPERATION_TYPE_BASE_MAX} respectively should be interpreted as an extension
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IBurst.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IBurst.aidl
index b089c49..a05a7fb 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/IBurst.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IBurst.aidl
@@ -17,6 +17,7 @@
 package android.hardware.neuralnetworks;
 
 import android.hardware.neuralnetworks.ErrorStatus;
+import android.hardware.neuralnetworks.ExecutionConfig;
 import android.hardware.neuralnetworks.ExecutionResult;
 import android.hardware.neuralnetworks.Request;
 
@@ -68,6 +69,8 @@
      *
      * Only a single execution on a given burst object may be active at any time.
      *
+     * Also see {@link IBurst::executeSynchronouslyWithConfig}.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed.
      * @param memoryIdentifierTokens A list of tokens where each token is a non-negative number
@@ -117,4 +120,13 @@
      *     - INVALID_ARGUMENT if one of the input arguments is invalid
      */
     void releaseMemoryResource(in long memoryIdentifierToken);
+
+    /**
+     * For detailed specification, please refer to {@link IBurst::executeSynchronously}. The
+     * difference between the two methods is that executeSynchronouslyWithConfig takes {@link
+     * ExecutionConfig} instead of a list of configuration parameters, and ExecutionConfig contains
+     * more configuration parameters than are passed to executeSynchronously.
+     */
+    ExecutionResult executeSynchronouslyWithConfig(in Request request,
+            in long[] memoryIdentifierTokens, in ExecutionConfig config, in long deadlineNs);
 }
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
index 72e2623..821b9fe 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl
@@ -28,6 +28,7 @@
 import android.hardware.neuralnetworks.IPreparedModelParcel;
 import android.hardware.neuralnetworks.Model;
 import android.hardware.neuralnetworks.NumberOfCacheFiles;
+import android.hardware.neuralnetworks.PrepareModelConfig;
 import android.hardware.neuralnetworks.Priority;
 
 /**
@@ -148,7 +149,7 @@
      *
      * If the device reports that caching is not supported, the user may avoid calling
      * IDevice::prepareModelFromCache or providing cache file descriptors to
-     * IDevice::prepareModel.
+     * IDevice::prepareModel or IDevice::prepareModelWithConfig.
      *
      * @return NumberOfCacheFiles structure indicating how many files for model and data cache the
      *     driver needs to cache a single prepared model. It must be less than or equal to
@@ -302,6 +303,8 @@
      *
      * Multiple threads may call prepareModel on the same model concurrently.
      *
+     * Also see {@link IDevice::prepareModelWithConfig}.
+     *
      * @param model The model to be prepared for execution.
      * @param preference Indicates the intended execution behavior of a prepared model.
      * @param priority The priority of the prepared model relative to other prepared models owned by
@@ -403,17 +406,17 @@
      * @param modelCache A vector of file descriptors for the security-sensitive cache. The length
      *                   of the vector must match the numModelCache returned from
      *                   getNumberOfCacheFilesNeeded. The cache file descriptors will be provided in
-     *                   the same order as with prepareModel.
+     *                   the same order as with prepareModel or prepareModelWithConfig.
      * @param dataCache A vector of file descriptors for the constants' cache. The length of the
      *                  vector must match the numDataCache returned from
      *                  getNumberOfCacheFilesNeeded. The cache file descriptors will be provided in
-     *                  the same order as with prepareModel.
+     *                  the same order as with prepareModel or prepareModelWithConfig.
      * @param token A caching token of length BYTE_SIZE_OF_CACHE_TOKEN identifying the prepared
      *              model. It is the same token provided when saving the cache files with
-     *              prepareModel. Tokens should be chosen to have a low rate of collision for a
-     *              particular application. The driver cannot detect a collision; a collision will
-     *              result in a failed execution or in a successful execution that produces
-     *              incorrect output values.
+     *              prepareModel or prepareModelWithConfig. Tokens should be chosen to have a low
+     *              rate of collision for a particular application. The driver cannot detect a
+     *              collision; a collision will result in a failed execution or in a successful
+     *              execution that produces incorrect output values.
      * @param callback A callback object used to return the error status of preparing the model for
      *                 execution and the prepared model if successful, nullptr otherwise. The
      *                 callback object's notify function must be called exactly once, even if the
@@ -429,4 +432,28 @@
     void prepareModelFromCache(in long deadlineNs, in ParcelFileDescriptor[] modelCache,
             in ParcelFileDescriptor[] dataCache, in byte[] token,
             in IPreparedModelCallback callback);
+
+    /**
+     * For detailed specification, please refer to {@link IDevice::prepareModel}. The only
+     * difference between the two methods is that prepareModelWithConfig takes {@link
+     * PrepareModelConfig} instead of standalone configuration parameters, which allows vendor
+     * specific compilation metadata to be passed.
+     *
+     * @param model The model to be prepared for execution.
+     * @param config Configuration parameters to prepare the model.
+     * @param callback A callback object used to return the error status of preparing the model for
+     *                 execution and the prepared model if successful, nullptr otherwise. The
+     *                 callback object's notify function must be called exactly once, even if the
+     *                 model could not be prepared.
+     * @throws ServiceSpecificException with one of the following ErrorStatus values:
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if one of the input arguments related to preparing the model is
+     *       invalid
+     *     - MISSED_DEADLINE_* if the preparation is aborted because the model cannot be prepared by
+     *       the deadline
+     *     - RESOURCE_EXHAUSTED_* if the task was aborted by the driver
+     */
+    void prepareModelWithConfig(
+            in Model model, in PrepareModelConfig config, in IPreparedModelCallback callback);
 }
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
index 79053e5..949804e 100644
--- a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl
@@ -18,6 +18,7 @@
 
 import android.hardware.common.NativeHandle;
 import android.hardware.neuralnetworks.ErrorStatus;
+import android.hardware.neuralnetworks.ExecutionConfig;
 import android.hardware.neuralnetworks.ExecutionResult;
 import android.hardware.neuralnetworks.FencedExecutionResult;
 import android.hardware.neuralnetworks.IBurst;
@@ -68,6 +69,8 @@
      * Any number of calls to the execute* functions, in any combination, may be made concurrently,
      * even on the same IPreparedModel object.
      *
+     * Also see {@link IPreparedModel::executeSynchronouslyWithConfig}.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed.
      * @param measure Specifies whether or not to measure duration of the execution. The duration
@@ -134,6 +137,8 @@
      * Any number of calls to the execute* functions, in any combination, may be made concurrently,
      * even on the same IPreparedModel object.
      *
+     * Also see {@link IPreparedModel::executeFencedWithConfig}.
+     *
      * @param request The input and output information on which the prepared model is to be
      *                executed. The outputs in the request must have fully specified dimensions.
      * @param waitFor A vector of sync fence file descriptors. Execution must not start until all
@@ -201,15 +206,7 @@
      *
      * @param request The input and output information on which the prepared model is to be
      *                executed.
-     * @param measure Specifies whether or not to measure duration of the execution.
-     * @param loopTimeoutDurationNs The maximum amount of time in nanoseconds that should be spent
-     *                              executing a {@link OperationType::WHILE} operation. If a loop
-     *                              condition model does not output false within this duration, the
-     *                              computation performed on the returned reusable execution object
-     *                              must be aborted. If -1 is provided, the maximum amount
-     *                              of time is {@link DEFAULT_LOOP_TIMEOUT_DURATION_NS}. Other
-     *                              negative values are invalid. When provided, the duration must
-     *                              not exceed {@link MAXIMUM_LOOP_TIMEOUT_DURATION_NS}.
+     * @param config Specifies the execution configuration parameters.
      * @return execution An IExecution object representing a reusable execution that has been
      *                   specialized for a fixed request.
      * @throws ServiceSpecificException with one of the following ErrorStatus values:
@@ -218,6 +215,64 @@
      *     - INVALID_ARGUMENT if one of the input arguments is invalid
      *     - RESOURCE_EXHAUSTED_* if the task was aborted by the driver
      */
-    IExecution createReusableExecution(
-            in Request request, in boolean measureTiming, in long loopTimeoutDurationNs);
+    IExecution createReusableExecution(in Request request, in ExecutionConfig config);
+
+    /**
+     * For detailed specification, please refer to {@link IPreparedModel::executeSynchronously}. The
+     * difference between the two methods is that executeSynchronouslyWithConfig takes {@link
+     * ExecutionConfig} instead of a list of configuration parameters, and ExecutionConfig contains
+     * more configuration parameters than are passed to executeSynchronously.
+     *
+     * @param request The input and output information on which the prepared model is to be
+     *                executed.
+     * @param config Specifies the execution configuration parameters.
+     * @param deadlineNs The time by which the execution is expected to complete. The time is
+     *                   measured in nanoseconds since boot (as from clock_gettime(CLOCK_BOOTTIME,
+     *                   &ts) or ::android::base::boot_clock). If the execution cannot be finished
+     *                   by the deadline, the execution may be aborted. Passing -1 means the
+     *                   deadline is omitted. Other negative valueggs are invalid.
+     * @return ExecutionResult parcelable, containing the status of the execution, output shapes and
+     *     timing information.
+     *     - MISSED_DEADLINE_* if the execution is aborted because it cannot be completed by the
+     *       deadline
+     *     - RESOURCE_EXHAUSTED_* if the task was aborted by the driver
+     */
+    ExecutionResult executeSynchronouslyWithConfig(
+            in Request request, in ExecutionConfig config, in long deadlineNs);
+
+    /**
+     * For detailed specification, please refer to {@link IPreparedModel::executeFenced}. The
+     * difference between the two methods is that executeFencedWithConfig takes {@link
+     * ExecutionConfig} instead of a list of configuration parameters, and ExecutionConfig contains
+     * more configuration parameters than are passed to executeFenced.
+     *
+     * @param request The input and output information on which the prepared model is to be
+     *                executed. The outputs in the request must have fully specified dimensions.
+     * @param waitFor A vector of sync fence file descriptors. Execution must not start until all
+     *                sync fences have been signaled.
+     * @param config Specifies the execution configuration parameters.
+     * @param deadlineNs The time by which the execution is expected to complete. The time is
+     *                   measured in nanoseconds since boot (as from clock_gettime(CLOCK_BOOTTIME,
+     *                   &ts) or ::android::base::boot_clock). If the execution cannot be finished
+     *                   by the deadline, the execution may be aborted. Passing -1 means the
+     *                   deadline is omitted. Other negative values are invalid.
+     * @param durationNs The length of time in nanoseconds within which the execution is expected to
+     *                   complete after all sync fences in waitFor are signaled. If the execution
+     *                   cannot be finished within the duration, the execution may be aborted.
+     *                   Passing -1 means the duration is omitted. Other negative values are
+     *                   invalid.
+     * @return The FencedExecutionResult parcelable, containing IFencedExecutionCallback and the
+     *         sync fence.
+     * @throws ServiceSpecificException with one of the following ErrorStatus values:
+     *     - DEVICE_UNAVAILABLE if driver is offline or busy
+     *     - GENERAL_FAILURE if there is an unspecified error
+     *     - INVALID_ARGUMENT if one of the input arguments is invalid, including fences in error
+     *       states.
+     *     - MISSED_DEADLINE_* if the execution is aborted because it cannot be completed by the
+     *       deadline
+     *     - RESOURCE_EXHAUSTED_* if the task was aborted by the driver
+     */
+    FencedExecutionResult executeFencedWithConfig(in Request request,
+            in ParcelFileDescriptor[] waitFor, in ExecutionConfig config, in long deadlineNs,
+            in long durationNs);
 }
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl
new file mode 100644
index 0000000..96df968
--- /dev/null
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/PrepareModelConfig.aidl
@@ -0,0 +1,95 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package android.hardware.neuralnetworks;
+
+import android.hardware.neuralnetworks.ExecutionPreference;
+import android.hardware.neuralnetworks.ExtensionNameAndPrefix;
+import android.hardware.neuralnetworks.Priority;
+import android.hardware.neuralnetworks.TokenValuePair;
+
+/**
+ * A type that is used to represent all configuration needed to
+ * prepare a model.
+ */
+@VintfStability
+parcelable PrepareModelConfig {
+    /**
+     * Indicates the intended execution behavior of a prepared model.
+     */
+    ExecutionPreference preference;
+    /**
+     * The priority of the prepared model relative to other prepared
+     * models owned by the client.
+     */
+    Priority priority;
+    /**
+     * The time by which the model is expected to be prepared. The
+     * time is measured in nanoseconds since boot (as from
+     * clock_gettime(CLOCK_BOOTTIME, &ts) or
+     * ::android::base::boot_clock). If the model cannot be prepared
+     * by the deadline, the preparation may be aborted. Passing -1
+     * means the deadline is omitted. Other negative values are
+     * invalid.
+     */
+    long deadlineNs;
+    /**
+     * A vector of file descriptors for the security-sensitive cache.
+     * The length of the vector must either be 0 indicating that
+     * caching information is not provided, or match the
+     * numModelCache returned from IDevice::getNumberOfCacheFilesNeeded. The
+     * cache file descriptors will be provided in the same order when
+     * retrieving the preparedModel from cache files with
+     * IDevice::prepareModelFromCache.
+     */
+    ParcelFileDescriptor[] modelCache;
+    /**
+     * A vector of file descriptors for the constants' cache. The
+     * length of the vector must either be 0 indicating that caching
+     * information is not provided, or match the numDataCache
+     * returned from IDevice::getNumberOfCacheFilesNeeded. The cache file
+     * descriptors will be provided in the same order when retrieving
+     * the preparedModel from cache files with IDevice::prepareModelFromCache.
+     */
+    ParcelFileDescriptor[] dataCache;
+    /**
+     * A caching token of length IDevice::BYTE_SIZE_OF_CACHE_TOKEN identifying
+     * the prepared model. The same token will be provided when
+     * retrieving the prepared model from the cache files with
+     * IDevice::prepareModelFromCache.  Tokens should be chosen to have a low
+     * rate of collision for a particular application. The driver
+     * cannot detect a collision; a collision will result in a failed
+     * execution or in a successful execution that produces incorrect
+     * output values. If both modelCache and dataCache are empty
+     * indicating that caching information is not provided, this
+     * token must be ignored.
+     */
+    byte[] cacheToken;
+    /**
+     * A vector of token / value pairs represent vendor specific
+     * compilation hints or metadata. The provided TokenValuePairs must not
+     * contain the same token twice. The driver must validate the
+     * data and ignore invalid hints. It is up to the driver to
+     * decide whether to respect the provided hints or not.
+     */
+    TokenValuePair[] compilationHints;
+    /**
+     * The mapping between extension names and prefixes of token values.
+     * The driver must ignore the corresponding compilation hint, if
+     * the extension is not supported.
+     */
+    ExtensionNameAndPrefix[] extensionNameToPrefix;
+}
diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/TokenValuePair.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/TokenValuePair.aidl
new file mode 100644
index 0000000..ec665b4
--- /dev/null
+++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/TokenValuePair.aidl
@@ -0,0 +1,42 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package android.hardware.neuralnetworks;
+
+/**
+ * A type that is used to represent a token / byte array data pair.
+ */
+@VintfStability
+parcelable TokenValuePair {
+    /**
+     * A 32bit integer token. The token is created by combining the
+     * extension prefix and enum defined within the extension.
+     * The low {@link IDevice::EXTENSION_TYPE_LOW_BITS_TYPE} bits of the value
+     * correspond to the hint within the extension and the high
+     * {@link IDevice::EXTENSION_TYPE_HIGH_BITS_PREFIX} bits encode the "prefix", which maps
+     * uniquely to the extension name. The sign bit is always 0.
+     *
+     * For example, if a token value is 0x7AAA000B and the corresponding
+     * {@link ExtensionNameAndPrefix} contains an entry with prefix=0x7AAA and
+     * name="vendor.test.test_extension", then the token should be interpreted as the hint
+     * 0x000B of the extension named vendor.test.test_extension.
+     */
+    int token;
+    /**
+     * A byte array containing the raw data.
+     */
+    byte[] value;
+}
diff --git a/neuralnetworks/aidl/utils/Android.bp b/neuralnetworks/aidl/utils/Android.bp
index 9437d5c..41eae85 100644
--- a/neuralnetworks/aidl/utils/Android.bp
+++ b/neuralnetworks/aidl/utils/Android.bp
@@ -111,19 +111,13 @@
     static_libs: [
         "libaidlcommonsupport",
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
         "libbase",
         "libbinder_ndk",
         "libcutils",
-        "libhidlbase",
-        "libhidlmemory",
-        "liblog",
-        "libutils",
     ],
     target: {
         android: {
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Burst.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Burst.h
index 0cc78d4..f2e6e75 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Burst.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Burst.h
@@ -86,10 +86,12 @@
                 GUARDED_BY(mMutex);
     };
 
+    // featureLevel is for testing purposes.
     static nn::GeneralResult<std::shared_ptr<const Burst>> create(
-            std::shared_ptr<aidl_hal::IBurst> burst);
+            std::shared_ptr<aidl_hal::IBurst> burst, nn::Version featureLevel);
 
-    Burst(PrivateConstructorTag tag, std::shared_ptr<aidl_hal::IBurst> burst);
+    Burst(PrivateConstructorTag tag, std::shared_ptr<aidl_hal::IBurst> burst,
+          nn::Version featureLevel);
 
     // See IBurst::cacheMemory for information.
     OptionalCacheHold cacheMemory(const nn::SharedMemory& memory) const override;
@@ -97,23 +99,29 @@
     // See IBurst::execute for information.
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     // See IBurst::createReusableExecution for information.
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> executeInternal(
             const aidl_hal::Request& request, const std::vector<int64_t>& memoryIdentifierTokens,
             bool measure, int64_t deadline, int64_t loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
             const hal::utils::RequestRelocation& relocation) const;
 
   private:
     mutable std::atomic_flag mExecutionInFlight = ATOMIC_FLAG_INIT;
     const std::shared_ptr<aidl_hal::IBurst> kBurst;
     const std::shared_ptr<MemoryCache> kMemoryCache;
+    const nn::Version kFeatureLevel;
 };
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
index 477b311..af58715 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
@@ -46,6 +46,10 @@
 #include <aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.h>
 #include <aidl/android/hardware/neuralnetworks/Timing.h>
 
+#ifdef NN_AIDL_V4_OR_ABOVE
+#include <aidl/android/hardware/neuralnetworks/TokenValuePair.h>
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 #include <android/binder_auto_utils.h>
 #include <nnapi/Result.h>
 #include <nnapi/Types.h>
@@ -74,7 +78,7 @@
         const aidl_hal::SymmPerChannelQuantParams& symmPerChannelQuantParams);
 GeneralResult<Operation> unvalidatedConvert(const aidl_hal::Operation& operation);
 GeneralResult<Model> unvalidatedConvert(const aidl_hal::Model& model);
-GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
+GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
         const aidl_hal::ExtensionNameAndPrefix& extensionNameAndPrefix);
 GeneralResult<Model::OperandValues> unvalidatedConvert(const std::vector<uint8_t>& operandValues);
 GeneralResult<Model::Subgraph> unvalidatedConvert(const aidl_hal::Subgraph& subgraph);
@@ -97,6 +101,10 @@
         const aidl_hal::ExtensionOperandTypeInformation& operandTypeInformation);
 GeneralResult<SharedHandle> unvalidatedConvert(const ndk::ScopedFileDescriptor& handle);
 
+#ifdef NN_AIDL_V4_OR_ABOVE
+GeneralResult<TokenValuePair> unvalidatedConvert(const aidl_hal::TokenValuePair& tokenValuePair);
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 GeneralResult<std::vector<Operation>> unvalidatedConvert(
         const std::vector<aidl_hal::Operation>& operations);
 
@@ -116,6 +124,14 @@
 
 GeneralResult<std::vector<Extension>> convert(const std::vector<aidl_hal::Extension>& extension);
 GeneralResult<std::vector<SharedMemory>> convert(const std::vector<aidl_hal::Memory>& memories);
+GeneralResult<std::vector<ExtensionNameAndPrefix>> convert(
+        const std::vector<aidl_hal::ExtensionNameAndPrefix>& extensionNameAndPrefix);
+
+#ifdef NN_AIDL_V4_OR_ABOVE
+GeneralResult<std::vector<TokenValuePair>> convert(
+        const std::vector<aidl_hal::TokenValuePair>& metaData);
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 GeneralResult<std::vector<OutputShape>> convert(
         const std::vector<aidl_hal::OutputShape>& outputShapes);
 GeneralResult<std::vector<SharedHandle>> convert(
@@ -152,7 +168,7 @@
 nn::GeneralResult<std::vector<uint8_t>> unvalidatedConvert(
         const nn::Model::OperandValues& operandValues);
 nn::GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
-        const nn::Model::ExtensionNameAndPrefix& extensionNameToPrefix);
+        const nn::ExtensionNameAndPrefix& extensionNameToPrefix);
 nn::GeneralResult<Model> unvalidatedConvert(const nn::Model& model);
 nn::GeneralResult<Priority> unvalidatedConvert(const nn::Priority& priority);
 nn::GeneralResult<Request> unvalidatedConvert(const nn::Request& request);
@@ -166,6 +182,10 @@
 nn::GeneralResult<Capabilities> unvalidatedConvert(const nn::Capabilities& capabilities);
 nn::GeneralResult<Extension> unvalidatedConvert(const nn::Extension& extension);
 
+#ifdef NN_AIDL_V4_OR_ABOVE
+nn::GeneralResult<TokenValuePair> unvalidatedConvert(const nn::TokenValuePair& tokenValuePair);
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 nn::GeneralResult<std::vector<uint8_t>> convert(const nn::CacheToken& cacheToken);
 nn::GeneralResult<BufferDesc> convert(const nn::BufferDesc& bufferDesc);
 nn::GeneralResult<DeviceType> convert(const nn::DeviceType& deviceType);
@@ -190,6 +210,13 @@
 nn::GeneralResult<std::vector<ndk::ScopedFileDescriptor>> convert(
         const std::vector<nn::SyncFence>& syncFences);
 nn::GeneralResult<std::vector<Extension>> convert(const std::vector<nn::Extension>& extensions);
+nn::GeneralResult<std::vector<ExtensionNameAndPrefix>> convert(
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix);
+
+#ifdef NN_AIDL_V4_OR_ABOVE
+nn::GeneralResult<std::vector<TokenValuePair>> convert(
+        const std::vector<nn::TokenValuePair>& metaData);
+#endif  // NN_AIDL_V4_OR_ABOVE
 
 nn::GeneralResult<std::vector<int32_t>> toSigned(const std::vector<uint32_t>& vec);
 
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Device.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Device.h
index d558f66..615c6de 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Device.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Device.h
@@ -42,6 +42,7 @@
     struct PrivateConstructorTag {};
 
   public:
+    // featureLevel is for testing purposes.
     static nn::GeneralResult<std::shared_ptr<const Device>> create(
             std::string name, std::shared_ptr<aidl_hal::IDevice> device, nn::Version featureLevel);
 
@@ -67,8 +68,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
index 205d428..cacdc26 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/HalInterfaces.h
@@ -63,7 +63,9 @@
 
 #ifdef NN_AIDL_V4_OR_ABOVE
 #include <aidl/android/hardware/neuralnetworks/BnExecution.h>
+#include <aidl/android/hardware/neuralnetworks/ExecutionConfig.h>
 #include <aidl/android/hardware/neuralnetworks/IExecution.h>
+#include <aidl/android/hardware/neuralnetworks/PrepareModelConfig.h>
 #endif  // NN_AIDL_V4_OR_ABOVE
 
 namespace android::nn {
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/InvalidDevice.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/InvalidDevice.h
index e66507a..9375c1d 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/InvalidDevice.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/InvalidDevice.h
@@ -53,6 +53,9 @@
             const std::vector<ndk::ScopedFileDescriptor>& dataCache,
             const std::vector<uint8_t>& token,
             const std::shared_ptr<IPreparedModelCallback>& callback) override;
+    ndk::ScopedAStatus prepareModelWithConfig(
+            const Model& model, const PrepareModelConfig& config,
+            const std::shared_ptr<IPreparedModelCallback>& callback) override;
     ndk::ScopedAStatus prepareModelFromCache(
             int64_t deadline, const std::vector<ndk::ScopedFileDescriptor>& modelCache,
             const std::vector<ndk::ScopedFileDescriptor>& dataCache,
diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
index 24cd681..cb6a85b 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/PreparedModel.h
@@ -40,6 +40,7 @@
     struct PrivateConstructorTag {};
 
   public:
+    // featureLevel is for testing purposes.
     static nn::GeneralResult<std::shared_ptr<const PreparedModel>> create(
             std::shared_ptr<aidl_hal::IPreparedModel> preparedModel, nn::Version featureLevel);
 
@@ -49,18 +50,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
@@ -68,6 +74,8 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> executeInternal(
             const Request& request, bool measure, int64_t deadline, int64_t loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
             const hal::utils::RequestRelocation& relocation) const;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
@@ -75,6 +83,8 @@
                           const std::vector<ndk::ScopedFileDescriptor>& waitFor, bool measure,
                           int64_t deadline, int64_t loopTimeoutDuration,
                           int64_t timeoutDurationAfterFence,
+                          const std::vector<nn::TokenValuePair>& hints,
+                          const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
                           const hal::utils::RequestRelocation& relocation) const;
 
   private:
diff --git a/neuralnetworks/aidl/utils/src/Burst.cpp b/neuralnetworks/aidl/utils/src/Burst.cpp
index fb00b26..6c7aa88 100644
--- a/neuralnetworks/aidl/utils/src/Burst.cpp
+++ b/neuralnetworks/aidl/utils/src/Burst.cpp
@@ -43,12 +43,16 @@
     static nn::GeneralResult<std::shared_ptr<const BurstExecution>> create(
             std::shared_ptr<const Burst> burst, Request request,
             std::vector<int64_t> memoryIdentifierTokens, bool measure, int64_t loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
             hal::utils::RequestRelocation relocation,
             std::vector<Burst::OptionalCacheHold> cacheHolds);
 
     BurstExecution(PrivateConstructorTag tag, std::shared_ptr<const Burst> burst, Request request,
                    std::vector<int64_t> memoryIdentifierTokens, bool measure,
-                   int64_t loopTimeoutDuration, hal::utils::RequestRelocation relocation,
+                   int64_t loopTimeoutDuration, const std::vector<nn::TokenValuePair>& hints,
+                   const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
+                   hal::utils::RequestRelocation relocation,
                    std::vector<Burst::OptionalCacheHold> cacheHolds);
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> compute(
@@ -64,6 +68,8 @@
     const std::vector<int64_t> kMemoryIdentifierTokens;
     const bool kMeasure;
     const int64_t kLoopTimeoutDuration;
+    const std::vector<nn::TokenValuePair> kHints;
+    const std::vector<nn::ExtensionNameAndPrefix> kExtensionNameToPrefix;
     const hal::utils::RequestRelocation kRelocation;
     const std::vector<Burst::OptionalCacheHold> kCacheHolds;
 };
@@ -149,17 +155,20 @@
 }
 
 nn::GeneralResult<std::shared_ptr<const Burst>> Burst::create(
-        std::shared_ptr<aidl_hal::IBurst> burst) {
+        std::shared_ptr<aidl_hal::IBurst> burst, nn::Version featureLevel) {
     if (burst == nullptr) {
         return NN_ERROR(nn::ErrorStatus::GENERAL_FAILURE)
                << "aidl_hal::utils::Burst::create must have non-null burst";
     }
 
-    return std::make_shared<const Burst>(PrivateConstructorTag{}, std::move(burst));
+    return std::make_shared<const Burst>(PrivateConstructorTag{}, std::move(burst), featureLevel);
 }
 
-Burst::Burst(PrivateConstructorTag /*tag*/, std::shared_ptr<aidl_hal::IBurst> burst)
-    : kBurst(std::move(burst)), kMemoryCache(std::make_shared<MemoryCache>(kBurst)) {
+Burst::Burst(PrivateConstructorTag /*tag*/, std::shared_ptr<aidl_hal::IBurst> burst,
+             nn::Version featureLevel)
+    : kBurst(std::move(burst)),
+      kMemoryCache(std::make_shared<MemoryCache>(kBurst)),
+      kFeatureLevel(featureLevel) {
     CHECK(kBurst != nullptr);
 }
 
@@ -170,8 +179,9 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -200,14 +210,14 @@
         memoryIdentifierTokens.push_back(-1);
     }
     CHECK_EQ(requestInShared.pools.size(), memoryIdentifierTokens.size());
-
     return executeInternal(aidlRequest, memoryIdentifierTokens, aidlMeasure, aidlDeadline,
-                           aidlLoopTimeoutDuration, relocation);
+                           aidlLoopTimeoutDuration, hints, extensionNameToPrefix, relocation);
 }
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> Burst::executeInternal(
         const Request& request, const std::vector<int64_t>& memoryIdentifierTokens, bool measure,
-        int64_t deadline, int64_t loopTimeoutDuration,
+        int64_t deadline, int64_t loopTimeoutDuration, const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
         const hal::utils::RequestRelocation& relocation) const {
     // Ensure that at most one execution is in flight at any given time.
     const bool alreadyInFlight = mExecutionInFlight.test_and_set();
@@ -221,9 +231,21 @@
     }
 
     ExecutionResult executionResult;
-    const auto ret = kBurst->executeSynchronously(request, memoryIdentifierTokens, measure,
-                                                  deadline, loopTimeoutDuration, &executionResult);
-    HANDLE_ASTATUS(ret) << "execute failed";
+    if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
+        auto aidlHints = NN_TRY(convert(hints));
+        auto aidlExtensionPrefix = NN_TRY(convert(extensionNameToPrefix));
+        const auto ret = kBurst->executeSynchronouslyWithConfig(
+                request, memoryIdentifierTokens,
+                {measure, loopTimeoutDuration, std::move(aidlHints),
+                 std::move(aidlExtensionPrefix)},
+                deadline, &executionResult);
+        HANDLE_ASTATUS(ret) << "execute failed";
+    } else {
+        const auto ret =
+                kBurst->executeSynchronously(request, memoryIdentifierTokens, measure, deadline,
+                                             loopTimeoutDuration, &executionResult);
+        HANDLE_ASTATUS(ret) << "execute failed";
+    }
     if (!executionResult.outputSufficientSize) {
         auto canonicalOutputShapes =
                 nn::convert(executionResult.outputShapes).value_or(std::vector<nn::OutputShape>{});
@@ -241,7 +263,9 @@
 
 nn::GeneralResult<nn::SharedExecution> Burst::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -272,12 +296,15 @@
 
     return BurstExecution::create(shared_from_this(), std::move(aidlRequest),
                                   std::move(memoryIdentifierTokens), aidlMeasure,
-                                  aidlLoopTimeoutDuration, std::move(relocation), std::move(holds));
+                                  aidlLoopTimeoutDuration, hints, extensionNameToPrefix,
+                                  std::move(relocation), std::move(holds));
 }
 
 nn::GeneralResult<std::shared_ptr<const BurstExecution>> BurstExecution::create(
         std::shared_ptr<const Burst> burst, Request request,
         std::vector<int64_t> memoryIdentifierTokens, bool measure, int64_t loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
         hal::utils::RequestRelocation relocation,
         std::vector<Burst::OptionalCacheHold> cacheHolds) {
     if (burst == nullptr) {
@@ -286,13 +313,15 @@
 
     return std::make_shared<const BurstExecution>(
             PrivateConstructorTag{}, std::move(burst), std::move(request),
-            std::move(memoryIdentifierTokens), measure, loopTimeoutDuration, std::move(relocation),
-            std::move(cacheHolds));
+            std::move(memoryIdentifierTokens), measure, loopTimeoutDuration, hints,
+            extensionNameToPrefix, std::move(relocation), std::move(cacheHolds));
 }
 
 BurstExecution::BurstExecution(PrivateConstructorTag /*tag*/, std::shared_ptr<const Burst> burst,
                                Request request, std::vector<int64_t> memoryIdentifierTokens,
                                bool measure, int64_t loopTimeoutDuration,
+                               const std::vector<nn::TokenValuePair>& hints,
+                               const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
                                hal::utils::RequestRelocation relocation,
                                std::vector<Burst::OptionalCacheHold> cacheHolds)
     : kBurst(std::move(burst)),
@@ -300,6 +329,8 @@
       kMemoryIdentifierTokens(std::move(memoryIdentifierTokens)),
       kMeasure(measure),
       kLoopTimeoutDuration(loopTimeoutDuration),
+      kHints(hints),
+      kExtensionNameToPrefix(extensionNameToPrefix),
       kRelocation(std::move(relocation)),
       kCacheHolds(std::move(cacheHolds)) {}
 
@@ -307,7 +338,8 @@
         const nn::OptionalTimePoint& deadline) const {
     const auto aidlDeadline = NN_TRY(convert(deadline));
     return kBurst->executeInternal(kRequest, kMemoryIdentifierTokens, kMeasure, aidlDeadline,
-                                   kLoopTimeoutDuration, kRelocation);
+                                   kLoopTimeoutDuration, kHints, kExtensionNameToPrefix,
+                                   kRelocation);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
diff --git a/neuralnetworks/aidl/utils/src/Conversions.cpp b/neuralnetworks/aidl/utils/src/Conversions.cpp
index 113d2da..eb28db7 100644
--- a/neuralnetworks/aidl/utils/src/Conversions.cpp
+++ b/neuralnetworks/aidl/utils/src/Conversions.cpp
@@ -302,9 +302,9 @@
     };
 }
 
-GeneralResult<Model::ExtensionNameAndPrefix> unvalidatedConvert(
+GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
         const aidl_hal::ExtensionNameAndPrefix& extensionNameAndPrefix) {
-    return Model::ExtensionNameAndPrefix{
+    return ExtensionNameAndPrefix{
             .name = extensionNameAndPrefix.name,
             .prefix = extensionNameAndPrefix.prefix,
     };
@@ -506,6 +506,12 @@
     return std::make_shared<const Handle>(std::move(duplicatedFd));
 }
 
+#ifdef NN_AIDL_V4_OR_ABOVE
+GeneralResult<TokenValuePair> unvalidatedConvert(const aidl_hal::TokenValuePair& tokenValuePair) {
+    return TokenValuePair{.token = tokenValuePair.token, .value = tokenValuePair.value};
+}
+#endif  // NN_AIDL_V4_OR_ABOVE
+
 GeneralResult<Capabilities> convert(const aidl_hal::Capabilities& capabilities) {
     return validatedConvert(capabilities);
 }
@@ -562,6 +568,17 @@
 GeneralResult<std::vector<SharedMemory>> convert(const std::vector<aidl_hal::Memory>& memories) {
     return validatedConvert(memories);
 }
+GeneralResult<std::vector<ExtensionNameAndPrefix>> convert(
+        const std::vector<aidl_hal::ExtensionNameAndPrefix>& extensionNameAndPrefix) {
+    return unvalidatedConvert(extensionNameAndPrefix);
+}
+
+#ifdef NN_AIDL_V4_OR_ABOVE
+GeneralResult<std::vector<TokenValuePair>> convert(
+        const std::vector<aidl_hal::TokenValuePair>& metaData) {
+    return validatedConvert(metaData);
+}
+#endif  // NN_AIDL_V4_OR_ABOVE
 
 GeneralResult<std::vector<OutputShape>> convert(
         const std::vector<aidl_hal::OutputShape>& outputShapes) {
@@ -942,7 +959,7 @@
 }
 
 nn::GeneralResult<ExtensionNameAndPrefix> unvalidatedConvert(
-        const nn::Model::ExtensionNameAndPrefix& extensionNameToPrefix) {
+        const nn::ExtensionNameAndPrefix& extensionNameToPrefix) {
     return ExtensionNameAndPrefix{
             .name = extensionNameToPrefix.name,
             .prefix = extensionNameToPrefix.prefix,
@@ -1055,6 +1072,11 @@
     return Extension{.name = extension.name,
                      .operandTypes = NN_TRY(unvalidatedConvert(extension.operandTypes))};
 }
+#ifdef NN_AIDL_V4_OR_ABOVE
+nn::GeneralResult<TokenValuePair> unvalidatedConvert(const nn::TokenValuePair& tokenValuePair) {
+    return TokenValuePair{.token = tokenValuePair.token, .value = tokenValuePair.value};
+}
+#endif  // NN_AIDL_V4_OR_ABOVE
 
 nn::GeneralResult<std::vector<uint8_t>> convert(const nn::CacheToken& cacheToken) {
     return validatedConvert(cacheToken);
@@ -1134,6 +1156,17 @@
         const std::vector<nn::SyncFence>& syncFences) {
     return validatedConvert(syncFences);
 }
+nn::GeneralResult<std::vector<ExtensionNameAndPrefix>> convert(
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) {
+    return unvalidatedConvert(extensionNameToPrefix);
+}
+
+#ifdef NN_AIDL_V4_OR_ABOVE
+nn::GeneralResult<std::vector<TokenValuePair>> convert(
+        const std::vector<nn::TokenValuePair>& metaData) {
+    return validatedConvert(metaData);
+}
+#endif  // NN_AIDL_V4_OR_ABOVE
 
 nn::GeneralResult<std::vector<Extension>> convert(const std::vector<nn::Extension>& extensions) {
     return validatedConvert(extensions);
diff --git a/neuralnetworks/aidl/utils/src/Device.cpp b/neuralnetworks/aidl/utils/src/Device.cpp
index bad10ed..f3f4fdb 100644
--- a/neuralnetworks/aidl/utils/src/Device.cpp
+++ b/neuralnetworks/aidl/utils/src/Device.cpp
@@ -215,7 +215,9 @@
 nn::GeneralResult<nn::SharedPreparedModel> Device::prepareModel(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
         nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that model is ready for IPC.
     std::optional<nn::Model> maybeModelInShared;
     const nn::Model& modelInShared =
@@ -225,17 +227,28 @@
     const auto aidlPreference = NN_TRY(convert(preference));
     const auto aidlPriority = NN_TRY(convert(priority));
     const auto aidlDeadline = NN_TRY(convert(deadline));
-    const auto aidlModelCache = NN_TRY(convert(modelCache));
-    const auto aidlDataCache = NN_TRY(convert(dataCache));
+    auto aidlModelCache = NN_TRY(convert(modelCache));
+    auto aidlDataCache = NN_TRY(convert(dataCache));
     const auto aidlToken = NN_TRY(convert(token));
 
     const auto cb = ndk::SharedRefBase::make<PreparedModelCallback>(kFeatureLevel);
     const auto scoped = kDeathHandler.protectCallback(cb.get());
 
+    if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
+        auto aidlHints = NN_TRY(convert(hints));
+        auto aidlExtensionPrefix = NN_TRY(convert(extensionNameToPrefix));
+        const auto ret = kDevice->prepareModelWithConfig(
+                aidlModel,
+                {aidlPreference, aidlPriority, aidlDeadline, std::move(aidlModelCache),
+                 std::move(aidlDataCache), aidlToken, std::move(aidlHints),
+                 std::move(aidlExtensionPrefix)},
+                cb);
+        HANDLE_ASTATUS(ret) << "prepareModel failed";
+        return cb->get();
+    }
     const auto ret = kDevice->prepareModel(aidlModel, aidlPreference, aidlPriority, aidlDeadline,
                                            aidlModelCache, aidlDataCache, aidlToken, cb);
     HANDLE_ASTATUS(ret) << "prepareModel failed";
-
     return cb->get();
 }
 
diff --git a/neuralnetworks/aidl/utils/src/Execution.cpp b/neuralnetworks/aidl/utils/src/Execution.cpp
index c4add63..2fd88af 100644
--- a/neuralnetworks/aidl/utils/src/Execution.cpp
+++ b/neuralnetworks/aidl/utils/src/Execution.cpp
@@ -63,7 +63,7 @@
 ExecutionWithCachedRequest::compute(const nn::OptionalTimePoint& deadline) const {
     const auto aidlDeadline = NN_TRY(convert(deadline));
     return kPreparedModel->executeInternal(kRequest, kMeasure, aidlDeadline, kLoopTimeoutDuration,
-                                           kRelocation);
+                                           {}, {}, kRelocation);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
@@ -73,9 +73,9 @@
     const auto aidlWaitFor = NN_TRY(convert(waitFor));
     const auto aidlDeadline = NN_TRY(convert(deadline));
     const auto aidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
-    return kPreparedModel->executeFencedInternal(kRequest, aidlWaitFor, kMeasure, aidlDeadline,
-                                                 kLoopTimeoutDuration,
-                                                 aidlTimeoutDurationAfterFence, kRelocation);
+    return kPreparedModel->executeFencedInternal(
+            kRequest, aidlWaitFor, kMeasure, aidlDeadline, kLoopTimeoutDuration,
+            aidlTimeoutDurationAfterFence, {}, {}, kRelocation);
 }
 
 nn::GeneralResult<std::shared_ptr<const Execution>> Execution::create(
diff --git a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
index c9d9955..33270ff 100644
--- a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
+++ b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
@@ -167,6 +167,31 @@
     return ndk::ScopedAStatus::ok();
 }
 
+ndk::ScopedAStatus InvalidDevice::prepareModelWithConfig(
+        const Model& model, const PrepareModelConfig& config,
+        const std::shared_ptr<IPreparedModelCallback>& callback) {
+    if (!utils::valid(config.extensionNameToPrefix)) {
+        callback->notify(ErrorStatus::INVALID_ARGUMENT, nullptr);
+        return toAStatus(ErrorStatus::INVALID_ARGUMENT, "Invalid extensionNameToPrefix");
+    }
+    for (const auto& hint : config.compilationHints) {
+        auto result = std::find_if(config.extensionNameToPrefix.begin(),
+                                   config.extensionNameToPrefix.end(),
+                                   [&hint](const ExtensionNameAndPrefix& extension) {
+                                       uint16_t prefix = static_cast<uint32_t>(hint.token) >>
+                                                         IDevice::EXTENSION_TYPE_LOW_BITS_TYPE;
+                                       return prefix == extension.prefix;
+                                   });
+        if (result == config.extensionNameToPrefix.end()) {
+            callback->notify(ErrorStatus::INVALID_ARGUMENT, nullptr);
+            return toAStatus(ErrorStatus::INVALID_ARGUMENT,
+                             "Invalid token for compilation hints: " + std::to_string(hint.token));
+        }
+    }
+    return prepareModel(model, config.preference, config.priority, config.deadlineNs,
+                        config.modelCache, config.dataCache, config.cacheToken, callback);
+}
+
 ndk::ScopedAStatus InvalidDevice::prepareModelFromCache(
         int64_t /*deadline*/, const std::vector<ndk::ScopedFileDescriptor>& /*modelCache*/,
         const std::vector<ndk::ScopedFileDescriptor>& /*dataCache*/,
diff --git a/neuralnetworks/aidl/utils/src/PreparedModel.cpp b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
index 6d1de56..7e3a31c 100644
--- a/neuralnetworks/aidl/utils/src/PreparedModel.cpp
+++ b/neuralnetworks/aidl/utils/src/PreparedModel.cpp
@@ -128,8 +128,9 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> PreparedModel::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -141,30 +142,46 @@
     const auto aidlMeasure = NN_TRY(convert(measure));
     const auto aidlDeadline = NN_TRY(convert(deadline));
     const auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
-    return executeInternal(aidlRequest, aidlMeasure, aidlDeadline, aidlLoopTimeoutDuration,
-                           relocation);
+    return executeInternal(aidlRequest, aidlMeasure, aidlDeadline, aidlLoopTimeoutDuration, hints,
+                           extensionNameToPrefix, relocation);
 }
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
 PreparedModel::executeInternal(const Request& request, bool measure, int64_t deadline,
                                int64_t loopTimeoutDuration,
+                               const std::vector<nn::TokenValuePair>& hints,
+                               const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
                                const hal::utils::RequestRelocation& relocation) const {
     if (relocation.input) {
         relocation.input->flush();
     }
 
     ExecutionResult executionResult;
-    const auto ret = kPreparedModel->executeSynchronously(request, measure, deadline,
-                                                          loopTimeoutDuration, &executionResult);
-    HANDLE_ASTATUS(ret) << "executeSynchronously failed";
+    if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
+        auto aidlHints = NN_TRY(convert(hints));
+        auto aidlExtensionPrefix = NN_TRY(convert(extensionNameToPrefix));
+        const auto ret = kPreparedModel->executeSynchronouslyWithConfig(
+                request,
+                {measure, loopTimeoutDuration, std::move(aidlHints),
+                 std::move(aidlExtensionPrefix)},
+                deadline, &executionResult);
+        HANDLE_ASTATUS(ret) << "executeSynchronouslyWithConfig failed";
+    } else {
+        const auto ret = kPreparedModel->executeSynchronously(
+                request, measure, deadline, loopTimeoutDuration, &executionResult);
+        HANDLE_ASTATUS(ret) << "executeSynchronously failed";
+    }
     return handleExecutionResult(executionResult, relocation);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFenced(const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
-                             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
-                             const nn::OptionalDuration& loopTimeoutDuration,
-                             const nn::OptionalDuration& timeoutDurationAfterFence) const {
+PreparedModel::executeFenced(
+        const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
+        nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const nn::OptionalDuration& timeoutDurationAfterFence,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -179,31 +196,45 @@
     const auto aidlLoopTimeoutDuration = NN_TRY(convert(loopTimeoutDuration));
     const auto aidlTimeoutDurationAfterFence = NN_TRY(convert(timeoutDurationAfterFence));
     return executeFencedInternal(aidlRequest, aidlWaitFor, aidlMeasure, aidlDeadline,
-                                 aidlLoopTimeoutDuration, aidlTimeoutDurationAfterFence,
-                                 relocation);
+                                 aidlLoopTimeoutDuration, aidlTimeoutDurationAfterFence, hints,
+                                 extensionNameToPrefix, relocation);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-PreparedModel::executeFencedInternal(const Request& request,
-                                     const std::vector<ndk::ScopedFileDescriptor>& waitFor,
-                                     bool measure, int64_t deadline, int64_t loopTimeoutDuration,
-                                     int64_t timeoutDurationAfterFence,
-                                     const hal::utils::RequestRelocation& relocation) const {
+PreparedModel::executeFencedInternal(
+        const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor, bool measure,
+        int64_t deadline, int64_t loopTimeoutDuration, int64_t timeoutDurationAfterFence,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix,
+        const hal::utils::RequestRelocation& relocation) const {
     if (relocation.input) {
         relocation.input->flush();
     }
 
     FencedExecutionResult result;
-    const auto ret =
-            kPreparedModel->executeFenced(request, waitFor, measure, deadline, loopTimeoutDuration,
-                                          timeoutDurationAfterFence, &result);
-    HANDLE_ASTATUS(ret) << "executeFenced failed";
+    if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
+        auto aidlHints = NN_TRY(convert(hints));
+        auto aidlExtensionPrefix = NN_TRY(convert(extensionNameToPrefix));
+        const auto ret = kPreparedModel->executeFencedWithConfig(
+                request, waitFor,
+                {measure, loopTimeoutDuration, std::move(aidlHints),
+                 std::move(aidlExtensionPrefix)},
+                deadline, timeoutDurationAfterFence, &result);
+        HANDLE_ASTATUS(ret) << "executeFencedWithConfig failed";
+    } else {
+        const auto ret = kPreparedModel->executeFenced(request, waitFor, measure, deadline,
+                                                       loopTimeoutDuration,
+                                                       timeoutDurationAfterFence, &result);
+        HANDLE_ASTATUS(ret) << "executeFenced failed";
+    }
     return handleFencedExecutionResult(result, relocation);
 }
 
 nn::GeneralResult<nn::SharedExecution> PreparedModel::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     // Ensure that request is ready for IPC.
     std::optional<nn::Request> maybeRequestInShared;
     hal::utils::RequestRelocation relocation;
@@ -217,8 +248,14 @@
 
     if (kFeatureLevel.level >= nn::Version::Level::FEATURE_LEVEL_8) {
         std::shared_ptr<IExecution> execution;
+        auto aidlHints = NN_TRY(convert(hints));
+        auto aidlExtensionPrefix = NN_TRY(convert(extensionNameToPrefix));
+
         const auto ret = kPreparedModel->createReusableExecution(
-                aidlRequest, aidlMeasure, aidlLoopTimeoutDuration, &execution);
+                aidlRequest,
+                {aidlMeasure, aidlLoopTimeoutDuration, std::move(aidlHints),
+                 std::move(aidlExtensionPrefix)},
+                &execution);
         HANDLE_ASTATUS(ret) << "createReusableExecution failed";
         return Execution::create(std::move(execution), std::move(relocation));
     }
@@ -232,7 +269,7 @@
     std::shared_ptr<IBurst> burst;
     const auto ret = kPreparedModel->configureExecutionBurst(&burst);
     HANDLE_ASTATUS(ret) << "configureExecutionBurst failed";
-    return Burst::create(std::move(burst));
+    return Burst::create(std::move(burst), kFeatureLevel);
 }
 
 std::any PreparedModel::getUnderlyingResource() const {
diff --git a/neuralnetworks/aidl/utils/test/DeviceTest.cpp b/neuralnetworks/aidl/utils/test/DeviceTest.cpp
index fb13af8..73727b3 100644
--- a/neuralnetworks/aidl/utils/test/DeviceTest.cpp
+++ b/neuralnetworks/aidl/utils/test/DeviceTest.cpp
@@ -61,7 +61,6 @@
                                                 .powerUsage = std::numeric_limits<float>::max()};
 constexpr NumberOfCacheFiles kNumberOfCacheFiles = {.numModelCache = nn::kMaxNumberOfCacheFiles - 1,
                                                     .numDataCache = nn::kMaxNumberOfCacheFiles};
-
 constexpr auto makeStatusOk = [] { return ndk::ScopedAStatus::ok(); };
 
 std::shared_ptr<MockDevice> createMockDevice() {
@@ -124,6 +123,18 @@
     };
 }
 
+const std::vector<nn::TokenValuePair> kHints = {nn::TokenValuePair{.token = 0, .value = {1}}};
+const std::vector<nn::ExtensionNameAndPrefix> kExtensionNameToPrefix = {
+        nn::ExtensionNameAndPrefix{.name = "com.android.nn_test", .prefix = 1}};
+auto makePreparedModelWithConfigReturn(ErrorStatus launchStatus, ErrorStatus returnStatus,
+                                       const std::shared_ptr<MockPreparedModel>& preparedModel) {
+    return [launchStatus, returnStatus, preparedModel](
+                   const Model& /*model*/, const PrepareModelConfig& /*config*/,
+                   const std::shared_ptr<IPreparedModelCallback>& cb) -> ndk::ScopedAStatus {
+        return makePreparedModelReturnImpl(launchStatus, returnStatus, preparedModel, cb);
+    };
+}
+
 auto makePreparedModelFromCacheReturn(ErrorStatus launchStatus, ErrorStatus returnStatus,
                                       const std::shared_ptr<MockPreparedModel>& preparedModel) {
     return [launchStatus, returnStatus, preparedModel](
@@ -560,6 +571,8 @@
 }
 
 TEST_P(DeviceTest, prepareModel) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -571,7 +584,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -580,6 +593,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelLaunchError) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -590,7 +605,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -598,6 +613,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelReturnError) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -608,7 +625,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -616,6 +633,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelNullptrError) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -626,7 +645,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -634,6 +653,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelTransportFailure) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -643,7 +664,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -651,6 +672,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelDeadObject) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup call
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -660,7 +683,7 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -668,6 +691,8 @@
 }
 
 TEST_P(DeviceTest, prepareModelAsyncCrash) {
+    if (kVersion.level > nn::Version::Level::FEATURE_LEVEL_7) return;
+
     // setup test
     const auto mockDevice = createMockDevice();
     const auto device = Device::create(kName, mockDevice, kVersion).value();
@@ -681,7 +706,157 @@
 
     // run test
     const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
-                                             nn::Priority::DEFAULT, {}, {}, {}, {});
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfig) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    const auto mockPreparedModel = MockPreparedModel::create();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makePreparedModelWithConfigReturn(ErrorStatus::NONE, ErrorStatus::NONE,
+                                                               mockPreparedModel)));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    EXPECT_NE(result.value(), nullptr);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigLaunchError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makePreparedModelWithConfigReturn(
+                    ErrorStatus::GENERAL_FAILURE, ErrorStatus::GENERAL_FAILURE, nullptr)));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigReturnError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makePreparedModelWithConfigReturn(
+                    ErrorStatus::NONE, ErrorStatus::GENERAL_FAILURE, nullptr)));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigNullptrError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makePreparedModelWithConfigReturn(ErrorStatus::NONE, ErrorStatus::NONE,
+                                                               nullptr)));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigTransportFailure) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigDeadObject) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST_P(DeviceTest, prepareModelWithConfigAsyncCrash) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockDevice = createMockDevice();
+    const auto device = Device::create(kName, mockDevice, kVersion).value();
+    const auto ret = [&device]() {
+        DeathMonitor::serviceDied(device->getDeathMonitor());
+        return ndk::ScopedAStatus::ok();
+    };
+    EXPECT_CALL(*mockDevice, prepareModelWithConfig(_, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(ret));
+
+    // run test
+    const auto result = device->prepareModel(kSimpleModel, nn::ExecutionPreference::DEFAULT,
+                                             nn::Priority::DEFAULT, {}, {}, {}, {}, kHints,
+                                             kExtensionNameToPrefix);
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/aidl/utils/test/MockBuffer.h b/neuralnetworks/aidl/utils/test/MockBuffer.h
index f77fa86..7a05a0f 100644
--- a/neuralnetworks/aidl/utils/test/MockBuffer.h
+++ b/neuralnetworks/aidl/utils/test/MockBuffer.h
@@ -21,7 +21,6 @@
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-#include <hidl/Status.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
diff --git a/neuralnetworks/aidl/utils/test/MockBurst.h b/neuralnetworks/aidl/utils/test/MockBurst.h
index 5083bbd..609bd30 100644
--- a/neuralnetworks/aidl/utils/test/MockBurst.h
+++ b/neuralnetworks/aidl/utils/test/MockBurst.h
@@ -21,7 +21,6 @@
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-#include <hidl/Status.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
@@ -32,6 +31,10 @@
                  bool measureTiming, int64_t deadline, int64_t loopTimeoutDuration,
                  ExecutionResult* executionResult),
                 (override));
+    MOCK_METHOD(ndk::ScopedAStatus, executeSynchronouslyWithConfig,
+                (const Request& request, const std::vector<int64_t>& memoryIdentifierTokens,
+                 const ExecutionConfig& config, int64_t deadline, ExecutionResult* executionResult),
+                (override));
     MOCK_METHOD(ndk::ScopedAStatus, releaseMemoryResource, (int64_t memoryIdentifierToken),
                 (override));
 };
diff --git a/neuralnetworks/aidl/utils/test/MockDevice.h b/neuralnetworks/aidl/utils/test/MockDevice.h
index 3a28d55..47b8346 100644
--- a/neuralnetworks/aidl/utils/test/MockDevice.h
+++ b/neuralnetworks/aidl/utils/test/MockDevice.h
@@ -50,6 +50,10 @@
                  const std::vector<uint8_t>& token,
                  const std::shared_ptr<IPreparedModelCallback>& callback),
                 (override));
+    MOCK_METHOD(ndk::ScopedAStatus, prepareModelWithConfig,
+                (const Model& model, const PrepareModelConfig& config,
+                 const std::shared_ptr<IPreparedModelCallback>& callback),
+                (override));
     MOCK_METHOD(ndk::ScopedAStatus, prepareModelFromCache,
                 (int64_t deadline, const std::vector<ndk::ScopedFileDescriptor>& modelCache,
                  const std::vector<ndk::ScopedFileDescriptor>& dataCache,
diff --git a/neuralnetworks/aidl/utils/test/MockExecution.h b/neuralnetworks/aidl/utils/test/MockExecution.h
index 216f569..782e54f 100644
--- a/neuralnetworks/aidl/utils/test/MockExecution.h
+++ b/neuralnetworks/aidl/utils/test/MockExecution.h
@@ -21,8 +21,6 @@
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-#include <hidl/HidlSupport.h>
-#include <hidl/Status.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
diff --git a/neuralnetworks/aidl/utils/test/MockFencedExecutionCallback.h b/neuralnetworks/aidl/utils/test/MockFencedExecutionCallback.h
index 06f9ea2..29449bb 100644
--- a/neuralnetworks/aidl/utils/test/MockFencedExecutionCallback.h
+++ b/neuralnetworks/aidl/utils/test/MockFencedExecutionCallback.h
@@ -22,7 +22,6 @@
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-#include <hidl/Status.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
diff --git a/neuralnetworks/aidl/utils/test/MockPreparedModel.h b/neuralnetworks/aidl/utils/test/MockPreparedModel.h
index 0ed9af9..a5b3b66 100644
--- a/neuralnetworks/aidl/utils/test/MockPreparedModel.h
+++ b/neuralnetworks/aidl/utils/test/MockPreparedModel.h
@@ -22,8 +22,6 @@
 #include <android/binder_interface_utils.h>
 #include <gmock/gmock.h>
 #include <gtest/gtest.h>
-#include <hidl/HidlSupport.h>
-#include <hidl/Status.h>
 
 namespace aidl::android::hardware::neuralnetworks::utils {
 
@@ -40,10 +38,19 @@
                  bool measureTiming, int64_t deadline, int64_t loopTimeoutDuration,
                  int64_t duration, FencedExecutionResult* fencedExecutionResult),
                 (override));
+    MOCK_METHOD(ndk::ScopedAStatus, executeSynchronouslyWithConfig,
+                (const Request& request, const ExecutionConfig& config, int64_t deadline,
+                 ExecutionResult* executionResult),
+                (override));
+    MOCK_METHOD(ndk::ScopedAStatus, executeFencedWithConfig,
+                (const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+                 const ExecutionConfig& config, int64_t deadline, int64_t duration,
+                 FencedExecutionResult* fencedExecutionResult),
+                (override));
     MOCK_METHOD(ndk::ScopedAStatus, configureExecutionBurst, (std::shared_ptr<IBurst> * burst),
                 (override));
     MOCK_METHOD(ndk::ScopedAStatus, createReusableExecution,
-                (const Request& request, bool measureTiming, int64_t loopTimeoutDuration,
+                (const Request& request, const ExecutionConfig& config,
                  std::shared_ptr<IExecution>* execution),
                 (override));
 };
diff --git a/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp b/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
index 8cfb7c1..bf6136d 100644
--- a/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
+++ b/neuralnetworks/aidl/utils/test/PreparedModelTest.cpp
@@ -70,6 +70,21 @@
 
 class PreparedModelTest : public VersionedAidlUtilsTestBase {};
 
+const std::vector<nn::TokenValuePair> kHints = {nn::TokenValuePair{.token = 0, .value = {1}}};
+const std::vector<nn::ExtensionNameAndPrefix> kExtensionNameToPrefix = {
+        nn::ExtensionNameAndPrefix{.name = "com.android.nn_test", .prefix = 1}};
+auto makeFencedExecutionWithConfigResult(
+        const std::shared_ptr<MockFencedExecutionCallback>& callback) {
+    return [callback](const Request& /*request*/,
+                      const std::vector<ndk::ScopedFileDescriptor>& /*waitFor*/,
+                      const ExecutionConfig& /*config*/, int64_t /*deadline*/, int64_t /*duration*/,
+                      FencedExecutionResult* fencedExecutionResult) {
+        *fencedExecutionResult = FencedExecutionResult{.callback = callback,
+                                                       .syncFence = ndk::ScopedFileDescriptor(-1)};
+        return ndk::ScopedAStatus::ok();
+    };
+}
+
 }  // namespace
 
 TEST_P(PreparedModelTest, invalidPreparedModel) {
@@ -82,6 +97,8 @@
 }
 
 TEST_P(PreparedModelTest, executeSync) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -96,7 +113,7 @@
                     DoAll(SetArgPointee<4>(mockExecutionResult), InvokeWithoutArgs(makeStatusOk)));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     EXPECT_TRUE(result.has_value())
@@ -104,6 +121,8 @@
 }
 
 TEST_P(PreparedModelTest, executeSyncError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -112,7 +131,7 @@
             .WillOnce(Invoke(makeGeneralFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -120,6 +139,8 @@
 }
 
 TEST_P(PreparedModelTest, executeSyncTransportFailure) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -128,7 +149,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -136,6 +157,8 @@
 }
 
 TEST_P(PreparedModelTest, executeSyncDeadObject) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -144,7 +167,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -152,6 +175,8 @@
 }
 
 TEST_P(PreparedModelTest, executeFenced) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -165,7 +190,7 @@
             .WillOnce(Invoke(makeFencedExecutionResult(mockCallback)));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -181,6 +206,8 @@
 }
 
 TEST_P(PreparedModelTest, executeFencedCallbackError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup call
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -195,7 +222,7 @@
             .WillOnce(Invoke(makeFencedExecutionResult(mockCallback)));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -211,6 +238,8 @@
 }
 
 TEST_P(PreparedModelTest, executeFencedError) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -219,7 +248,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -227,6 +256,8 @@
 }
 
 TEST_P(PreparedModelTest, executeFencedTransportFailure) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -235,7 +266,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -243,6 +274,8 @@
 }
 
 TEST_P(PreparedModelTest, executeFencedDeadObject) {
+    if (kVersion.level >= nn::Version::Level::FEATURE_LEVEL_8) return;
+
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
@@ -251,7 +284,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -276,7 +309,7 @@
                     DoAll(SetArgPointee<4>(mockExecutionResult), InvokeWithoutArgs(makeStatusOk)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -300,7 +333,7 @@
             .WillOnce(Invoke(makeGeneralFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -322,7 +355,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -344,7 +377,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -372,7 +405,7 @@
             .WillRepeatedly(Invoke(makeFencedExecutionResult(mockCallback)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -410,7 +443,7 @@
             .WillOnce(Invoke(makeFencedExecutionResult(mockCallback)));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -440,7 +473,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -462,7 +495,7 @@
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -484,7 +517,7 @@
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
 
     // create execution
-    const auto createResult = preparedModel->createReusableExecution({}, {}, {});
+    const auto createResult = preparedModel->createReusableExecution({}, {}, {}, {}, {});
     ASSERT_TRUE(createResult.has_value())
             << "Failed with " << createResult.error().code << ": " << createResult.error().message;
     ASSERT_NE(createResult.value(), nullptr);
@@ -495,6 +528,206 @@
     EXPECT_EQ(computeResult.error().code, nn::ErrorStatus::DEAD_OBJECT);
 }
 
+TEST_P(PreparedModelTest, executeSyncWithConfig) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    const auto mockExecutionResult = ExecutionResult{
+            .outputSufficientSize = true,
+            .outputShapes = {},
+            .timing = kNoTiming,
+    };
+    EXPECT_CALL(*mockPreparedModel, executeSynchronouslyWithConfig(_, _, _, _))
+            .Times(1)
+            .WillOnce(
+                    DoAll(SetArgPointee<3>(mockExecutionResult), InvokeWithoutArgs(makeStatusOk)));
+
+    // run test
+    const auto result = preparedModel->execute({}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    EXPECT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+}
+
+TEST_P(PreparedModelTest, executeSyncWithConfigError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeSynchronouslyWithConfig(_, _, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makeGeneralFailure));
+
+    // run test
+    const auto result = preparedModel->execute({}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, executeSyncWithConfigTransportFailure) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeSynchronouslyWithConfig(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+    // run test
+    const auto result = preparedModel->execute({}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, executeSyncWithConfigDeadObject) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeSynchronouslyWithConfig(_, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+    // run test
+    const auto result = preparedModel->execute({}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
+TEST_P(PreparedModelTest, executeFencedWithConfig) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    const auto mockCallback = MockFencedExecutionCallback::create();
+    EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
+            .Times(1)
+            .WillOnce(DoAll(SetArgPointee<0>(kNoTiming), SetArgPointee<1>(kNoTiming),
+                            SetArgPointee<2>(ErrorStatus::NONE), Invoke(makeStatusOk)));
+    EXPECT_CALL(*mockPreparedModel, executeFencedWithConfig(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makeFencedExecutionWithConfigResult(mockCallback)));
+
+    // run test
+    const auto result =
+            preparedModel->executeFenced({}, {}, {}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    const auto& [syncFence, callback] = result.value();
+    EXPECT_EQ(syncFence.syncWait({}), nn::SyncFence::FenceState::SIGNALED);
+    ASSERT_NE(callback, nullptr);
+
+    // get results from callback
+    const auto callbackResult = callback();
+    ASSERT_TRUE(callbackResult.has_value()) << "Failed with " << callbackResult.error().code << ": "
+                                            << callbackResult.error().message;
+}
+
+TEST_P(PreparedModelTest, executeFencedWithConfigCallbackError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup call
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    const auto mockCallback = MockFencedExecutionCallback::create();
+    EXPECT_CALL(*mockCallback, getExecutionInfo(_, _, _))
+            .Times(1)
+            .WillOnce(Invoke(DoAll(SetArgPointee<0>(kNoTiming), SetArgPointee<1>(kNoTiming),
+                                   SetArgPointee<2>(ErrorStatus::GENERAL_FAILURE),
+                                   Invoke(makeStatusOk))));
+    EXPECT_CALL(*mockPreparedModel, executeFencedWithConfig(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(Invoke(makeFencedExecutionWithConfigResult(mockCallback)));
+
+    // run test
+    const auto result =
+            preparedModel->executeFenced({}, {}, {}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_TRUE(result.has_value())
+            << "Failed with " << result.error().code << ": " << result.error().message;
+    const auto& [syncFence, callback] = result.value();
+    EXPECT_NE(syncFence.syncWait({}), nn::SyncFence::FenceState::ACTIVE);
+    ASSERT_NE(callback, nullptr);
+
+    // verify callback failure
+    const auto callbackResult = callback();
+    ASSERT_FALSE(callbackResult.has_value());
+    EXPECT_EQ(callbackResult.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, executeFencedWithConfigError) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeFencedWithConfig(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
+
+    // run test
+    const auto result =
+            preparedModel->executeFenced({}, {}, {}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, executeFencedWithConfigTransportFailure) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeFencedWithConfig(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
+
+    // run test
+    const auto result =
+            preparedModel->executeFenced({}, {}, {}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::GENERAL_FAILURE);
+}
+
+TEST_P(PreparedModelTest, executeFencedWithConfigDeadObject) {
+    if (kVersion.level < nn::Version::Level::FEATURE_LEVEL_8) return;
+
+    // setup test
+    const auto mockPreparedModel = MockPreparedModel::create();
+    const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
+    EXPECT_CALL(*mockPreparedModel, executeFencedWithConfig(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
+
+    // run test
+    const auto result =
+            preparedModel->executeFenced({}, {}, {}, {}, {}, {}, kHints, kExtensionNameToPrefix);
+
+    // verify result
+    ASSERT_FALSE(result.has_value());
+    EXPECT_EQ(result.error().code, nn::ErrorStatus::DEAD_OBJECT);
+}
+
 TEST_P(PreparedModelTest, configureExecutionBurst) {
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
@@ -567,13 +800,13 @@
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
     const auto mockExecution = ndk::SharedRefBase::make<MockExecution>();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
             .Times(1)
-            .WillOnce(DoAll(SetArgPointee<3>(mockExecution), Invoke(makeStatusOk)));
+            .WillOnce(DoAll(SetArgPointee<2>(mockExecution), Invoke(makeStatusOk)));
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -586,13 +819,13 @@
 
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralFailure));
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -604,13 +837,13 @@
 
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeGeneralTransportFailure));
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -622,13 +855,13 @@
 
     // setup test
     const auto mockPreparedModel = MockPreparedModel::create();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
             .Times(1)
             .WillOnce(InvokeWithoutArgs(makeDeadObjectFailure));
     const auto preparedModel = PreparedModel::create(mockPreparedModel, kVersion).value();
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
index 2460fba..8c8a87a 100644
--- a/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
+++ b/neuralnetworks/aidl/vts/functional/GeneratedTestHarness.cpp
@@ -63,6 +63,8 @@
     // it is skipped. The field is set to true by default and is set to false in
     // quantization coupling tests to suppress skipping a test
     bool reportSkipping;
+    // `useConfig` indicates if a test should use execute*WithConfig functions for the execution.
+    bool useConfig;
     TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
                bool reusable)
         : executor(executor),
@@ -70,7 +72,8 @@
           outputType(outputType),
           memoryType(memoryType),
           reusable(reusable),
-          reportSkipping(true) {}
+          reportSkipping(true),
+          useConfig(false) {}
     TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
                bool reusable, bool reportSkipping)
         : executor(executor),
@@ -78,7 +81,17 @@
           outputType(outputType),
           memoryType(memoryType),
           reusable(reusable),
-          reportSkipping(reportSkipping) {}
+          reportSkipping(reportSkipping),
+          useConfig(false) {}
+    TestConfig(Executor executor, bool measureTiming, OutputType outputType, MemoryType memoryType,
+               bool reusable, bool reportSkipping, bool useConfig)
+        : executor(executor),
+          measureTiming(measureTiming),
+          outputType(outputType),
+          memoryType(memoryType),
+          reusable(reusable),
+          reportSkipping(reportSkipping),
+          useConfig(useConfig) {}
 };
 
 std::string toString(OutputType type) {
@@ -100,7 +113,8 @@
        << ", .measureTiming=" << (config.measureTiming ? "true" : "false")
        << ", .outputType=" << toString(config.outputType)
        << ", .memoryType=" << toString(config.memoryType)
-       << ", .reusable=" << (config.reusable ? "true" : "false") << "}";
+       << ", .reusable=" << (config.reusable ? "true" : "false")
+       << ", .useConfig=" << (config.useConfig ? "true" : "false") << "}";
     return ss.str();
 }
 
@@ -587,8 +601,8 @@
 
     std::shared_ptr<IExecution> execution;
     if (testConfig.reusable) {
-        const auto ret = preparedModel->createReusableExecution(request, testConfig.measureTiming,
-                                                                loopTimeoutDurationNs, &execution);
+        const auto ret = preparedModel->createReusableExecution(
+                request, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}}, &execution);
         ASSERT_TRUE(ret.isOk()) << static_cast<nn::ErrorStatus>(ret.getServiceSpecificError());
         ASSERT_NE(nullptr, execution.get());
     }
@@ -607,6 +621,10 @@
                 ::ndk::ScopedAStatus ret;
                 if (testConfig.reusable) {
                     ret = execution->executeSynchronously(kNoDeadline, &executionResult);
+                } else if (testConfig.useConfig) {
+                    ret = preparedModel->executeSynchronouslyWithConfig(
+                            request, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+                            kNoDeadline, &executionResult);
                 } else {
                     ret = preparedModel->executeSynchronously(request, testConfig.measureTiming,
                                                               kNoDeadline, loopTimeoutDurationNs,
@@ -649,9 +667,16 @@
 
                 ExecutionResult executionResult;
                 // execute
-                ret = burst->executeSynchronously(request, slots, testConfig.measureTiming,
-                                                  kNoDeadline, loopTimeoutDurationNs,
-                                                  &executionResult);
+                if (testConfig.useConfig) {
+                    ret = burst->executeSynchronouslyWithConfig(
+                            request, slots,
+                            {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}}, kNoDeadline,
+                            &executionResult);
+                } else {
+                    ret = burst->executeSynchronously(request, slots, testConfig.measureTiming,
+                                                      kNoDeadline, loopTimeoutDurationNs,
+                                                      &executionResult);
+                }
                 ASSERT_TRUE(ret.isOk() || ret.getExceptionCode() == EX_SERVICE_SPECIFIC)
                         << ret.getDescription();
                 if (ret.isOk()) {
@@ -680,6 +705,10 @@
                 ::ndk::ScopedAStatus ret;
                 if (testConfig.reusable) {
                     ret = execution->executeFenced({}, kNoDeadline, kNoDuration, &executionResult);
+                } else if (testConfig.useConfig) {
+                    ret = preparedModel->executeFencedWithConfig(
+                            request, {}, {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+                            kNoDeadline, kNoDuration, &executionResult);
                 } else {
                     ret = preparedModel->executeFenced(request, {}, testConfig.measureTiming,
                                                        kNoDeadline, loopTimeoutDurationNs,
@@ -697,9 +726,19 @@
                     waitFor.emplace_back(dupFd);
                     // If a sync fence is returned, try start another run waiting for the sync
                     // fence.
-                    ret = preparedModel->executeFenced(request, waitFor, testConfig.measureTiming,
-                                                       kNoDeadline, loopTimeoutDurationNs,
-                                                       kNoDuration, &executionResult);
+                    if (testConfig.reusable) {
+                        ret = execution->executeFenced(waitFor, kNoDeadline, kNoDuration,
+                                                       &executionResult);
+                    } else if (testConfig.useConfig) {
+                        ret = preparedModel->executeFencedWithConfig(
+                                request, waitFor,
+                                {testConfig.measureTiming, loopTimeoutDurationNs, {}, {}},
+                                kNoDeadline, kNoDuration, &executionResult);
+                    } else {
+                        ret = preparedModel->executeFenced(
+                                request, waitFor, testConfig.measureTiming, kNoDeadline,
+                                loopTimeoutDurationNs, kNoDuration, &executionResult);
+                    }
                     ASSERT_TRUE(ret.isOk());
                     waitForSyncFence(executionResult.syncFence.get());
                 }
@@ -830,11 +869,13 @@
     std::vector<Executor> executorList;
     std::vector<MemoryType> memoryTypeList;
     std::vector<bool> reusableList = {false};
+    std::vector<bool> useConfigList = {false};
 
     int deviceVersion;
     ASSERT_TRUE(device->getInterfaceVersion(&deviceVersion).isOk());
     if (deviceVersion >= kMinAidlLevelForFL8) {
         reusableList.push_back(true);
+        useConfigList.push_back(true);
     }
 
     switch (testKind) {
@@ -879,11 +920,14 @@
             for (const Executor executor : executorList) {
                 for (const MemoryType memoryType : memoryTypeList) {
                     for (const bool reusable : reusableList) {
-                        if (executor == Executor::BURST && reusable) continue;
-                        const TestConfig testConfig(executor, measureTiming, outputType, memoryType,
-                                                    reusable);
-                        SCOPED_TRACE(toString(testConfig));
-                        EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+                        for (const bool useConfig : useConfigList) {
+                            if ((useConfig || executor == Executor::BURST) && reusable) continue;
+                            const TestConfig testConfig(executor, measureTiming, outputType,
+                                                        memoryType, reusable,
+                                                        /*reportSkipping=*/true, useConfig);
+                            SCOPED_TRACE(toString(testConfig));
+                            EvaluatePreparedModel(device, preparedModel, testModel, testConfig);
+                        }
                     }
                 }
             }
@@ -942,6 +986,13 @@
             createPreparedModel(device, model, &preparedModel);
             if (preparedModel == nullptr) return;
             EvaluatePreparedModel(device, preparedModel, testModel, testKind);
+            int32_t deviceVersion;
+            ASSERT_TRUE(device->getInterfaceVersion(&deviceVersion).isOk());
+            if (deviceVersion >= kMinAidlLevelForFL8) {
+                createPreparedModel(device, model, &preparedModel, /*reportSkipping*/ true,
+                                    /*useConfig*/ true);
+                EvaluatePreparedModel(device, preparedModel, testModel, testKind);
+            }
         } break;
         case TestKind::QUANTIZATION_COUPLING: {
             ASSERT_TRUE(testModel.hasQuant8CoupledOperands());
diff --git a/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
index b3e9c63..97760ae 100644
--- a/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
+++ b/neuralnetworks/aidl/vts/functional/MemoryDomainTests.cpp
@@ -204,11 +204,23 @@
         return ndk::ScopedAStatus::fromServiceSpecificError(
                 static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
     }
+    ndk::ScopedAStatus executeSynchronouslyWithConfig(const Request&, const ExecutionConfig&,
+                                                      int64_t, ExecutionResult*) override {
+        return ndk::ScopedAStatus::fromServiceSpecificError(
+                static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+    }
+    ndk::ScopedAStatus executeFencedWithConfig(const Request&,
+                                               const std::vector<ndk::ScopedFileDescriptor>&,
+                                               const ExecutionConfig&, int64_t, int64_t,
+                                               FencedExecutionResult*) override {
+        return ndk::ScopedAStatus::fromServiceSpecificError(
+                static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
+    }
     ndk::ScopedAStatus configureExecutionBurst(std::shared_ptr<IBurst>*) override {
         return ndk::ScopedAStatus::fromServiceSpecificError(
                 static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
     }
-    ndk::ScopedAStatus createReusableExecution(const aidl_hal::Request&, bool, int64_t,
+    ndk::ScopedAStatus createReusableExecution(const aidl_hal::Request&, const ExecutionConfig&,
                                                std::shared_ptr<aidl_hal::IExecution>*) override {
         return ndk::ScopedAStatus::fromServiceSpecificError(
                 static_cast<int32_t>(ErrorStatus::GENERAL_FAILURE));
diff --git a/neuralnetworks/aidl/vts/functional/ValidateModel.cpp b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
index fdc7eff..931ba25 100644
--- a/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
+++ b/neuralnetworks/aidl/vts/functional/ValidateModel.cpp
@@ -77,6 +77,28 @@
     ASSERT_EQ(nullptr, preparedModel.get());
 }
 
+static void validatePrepareModelWithConfig(const std::shared_ptr<IDevice>& device,
+                                           const std::string& message, const Model& model,
+                                           ExecutionPreference preference, Priority priority) {
+    SCOPED_TRACE(message + " [prepareModelWithConfig]");
+
+    std::shared_ptr<PreparedModelCallback> preparedModelCallback =
+            ndk::SharedRefBase::make<PreparedModelCallback>();
+    const auto prepareLaunchStatus = device->prepareModelWithConfig(
+            model, {preference, priority, kNoDeadline, {}, {}, kEmptyCacheToken, {}, {}},
+            preparedModelCallback);
+    ASSERT_FALSE(prepareLaunchStatus.isOk());
+    ASSERT_EQ(prepareLaunchStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+    ASSERT_EQ(static_cast<ErrorStatus>(prepareLaunchStatus.getServiceSpecificError()),
+              ErrorStatus::INVALID_ARGUMENT);
+
+    preparedModelCallback->wait();
+    ErrorStatus prepareReturnStatus = preparedModelCallback->getStatus();
+    ASSERT_EQ(ErrorStatus::INVALID_ARGUMENT, prepareReturnStatus);
+    std::shared_ptr<IPreparedModel> preparedModel = preparedModelCallback->getPreparedModel();
+    ASSERT_EQ(nullptr, preparedModel.get());
+}
+
 static bool validExecutionPreference(ExecutionPreference preference) {
     return preference == ExecutionPreference::LOW_POWER ||
            preference == ExecutionPreference::FAST_SINGLE_ANSWER ||
@@ -103,6 +125,13 @@
     }
 
     validatePrepareModel(device, message, model, preference, priority);
+
+    int32_t aidlVersion;
+    ASSERT_TRUE(device->getInterfaceVersion(&aidlVersion).isOk());
+    if (aidlVersion >= kMinAidlLevelForFL8) {
+        // prepareModelWithConfig must satisfy all requirements enforced by prepareModel.
+        validatePrepareModelWithConfig(device, message, model, preference, priority);
+    }
 }
 
 static uint32_t addOperand(Model* model) {
diff --git a/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp b/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp
index e8debf7..d749841 100644
--- a/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp
+++ b/neuralnetworks/aidl/vts/functional/ValidateRequest.cpp
@@ -45,7 +45,7 @@
     {
         SCOPED_TRACE(message + " [createReusableExecution]");
         const auto createStatus = preparedModel->createReusableExecution(
-                request, measure, kOmittedTimeoutDuration, &execution);
+                request, {measure, kOmittedTimeoutDuration, {}, {}}, &execution);
         if (!createStatus.isOk()) {
             ASSERT_EQ(createStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
             ASSERT_EQ(static_cast<ErrorStatus>(createStatus.getServiceSpecificError()),
@@ -149,10 +149,59 @@
 
     int32_t aidlVersion;
     ASSERT_TRUE(preparedModel->getInterfaceVersion(&aidlVersion).isOk());
+    if (aidlVersion < kMinAidlLevelForFL8) {
+        return;
+    }
 
     // validate reusable execution
-    if (aidlVersion >= kMinAidlLevelForFL8) {
-        validateReusableExecution(preparedModel, message, request, measure);
+    validateReusableExecution(preparedModel, message, request, measure);
+
+    // synchronous with empty hints
+    {
+        SCOPED_TRACE(message + " [executeSynchronouslyWithConfig]");
+        ExecutionResult executionResult;
+        const auto executeStatus = preparedModel->executeSynchronouslyWithConfig(
+                request, {measure, kOmittedTimeoutDuration, {}, {}}, kNoDeadline, &executionResult);
+        ASSERT_FALSE(executeStatus.isOk());
+        ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_EQ(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()),
+                  ErrorStatus::INVALID_ARGUMENT);
+    }
+
+    // fenced with empty hints
+    {
+        SCOPED_TRACE(message + " [executeFencedWithConfig]");
+        FencedExecutionResult executionResult;
+        const auto executeStatus = preparedModel->executeFencedWithConfig(
+                request, {}, {false, kOmittedTimeoutDuration, {}, {}}, kNoDeadline, kNoDuration,
+                &executionResult);
+        ASSERT_FALSE(executeStatus.isOk());
+        ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_EQ(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()),
+                  ErrorStatus::INVALID_ARGUMENT);
+    }
+
+    // burst with empty hints
+    {
+        SCOPED_TRACE(message + " [burst executeSynchronouslyWithConfig]");
+
+        // create burst
+        std::shared_ptr<IBurst> burst;
+        auto ret = preparedModel->configureExecutionBurst(&burst);
+        ASSERT_TRUE(ret.isOk()) << ret.getDescription();
+        ASSERT_NE(nullptr, burst.get());
+
+        // use -1 for all memory identifier tokens
+        const std::vector<int64_t> slots(request.pools.size(), -1);
+
+        ExecutionResult executionResult;
+        const auto executeStatus = burst->executeSynchronouslyWithConfig(
+                request, slots, {measure, kOmittedTimeoutDuration, {}, {}}, kNoDeadline,
+                &executionResult);
+        ASSERT_FALSE(executeStatus.isOk());
+        ASSERT_EQ(executeStatus.getExceptionCode(), EX_SERVICE_SPECIFIC);
+        ASSERT_EQ(static_cast<ErrorStatus>(executeStatus.getServiceSpecificError()),
+                  ErrorStatus::INVALID_ARGUMENT);
     }
 }
 
diff --git a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
index c417356..ad93e6d 100644
--- a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
+++ b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.cpp
@@ -41,7 +41,8 @@
 
 // internal helper function
 void createPreparedModel(const std::shared_ptr<IDevice>& device, const Model& model,
-                         std::shared_ptr<IPreparedModel>* preparedModel, bool reportSkipping) {
+                         std::shared_ptr<IPreparedModel>* preparedModel, bool reportSkipping,
+                         bool useConfig) {
     ASSERT_NE(nullptr, preparedModel);
     *preparedModel = nullptr;
 
@@ -56,11 +57,25 @@
     // launch prepare model
     const std::shared_ptr<PreparedModelCallback> preparedModelCallback =
             ndk::SharedRefBase::make<PreparedModelCallback>();
-    const auto prepareLaunchStatus =
-            device->prepareModel(model, ExecutionPreference::FAST_SINGLE_ANSWER, kDefaultPriority,
-                                 kNoDeadline, {}, {}, kEmptyCacheToken, preparedModelCallback);
-    ASSERT_TRUE(prepareLaunchStatus.isOk()) << prepareLaunchStatus.getDescription();
-
+    if (useConfig) {
+        const auto prepareLaunchStatus =
+                device->prepareModelWithConfig(model,
+                                               {ExecutionPreference::FAST_SINGLE_ANSWER,
+                                                kDefaultPriority,
+                                                kNoDeadline,
+                                                {},
+                                                {},
+                                                kEmptyCacheToken,
+                                                {},
+                                                {}},
+                                               preparedModelCallback);
+        ASSERT_TRUE(prepareLaunchStatus.isOk()) << prepareLaunchStatus.getDescription();
+    } else {
+        const auto prepareLaunchStatus = device->prepareModel(
+                model, ExecutionPreference::FAST_SINGLE_ANSWER, kDefaultPriority, kNoDeadline, {},
+                {}, kEmptyCacheToken, preparedModelCallback);
+        ASSERT_TRUE(prepareLaunchStatus.isOk()) << prepareLaunchStatus.getDescription();
+    }
     // retrieve prepared model
     preparedModelCallback->wait();
     const ErrorStatus prepareReturnStatus = preparedModelCallback->getStatus();
diff --git a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h
index a900590..00d705c 100644
--- a/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h
+++ b/neuralnetworks/aidl/vts/functional/VtsHalNeuralnetworks.h
@@ -51,8 +51,8 @@
 // Create an IPreparedModel object. If the model cannot be prepared,
 // "preparedModel" will be nullptr instead.
 void createPreparedModel(const std::shared_ptr<IDevice>& device, const Model& model,
-                         std::shared_ptr<IPreparedModel>* preparedModel,
-                         bool reportSkipping = true);
+                         std::shared_ptr<IPreparedModel>* preparedModel, bool reportSkipping = true,
+                         bool useConfig = false);
 
 enum class Executor { SYNC, BURST, FENCED };
 
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h
index f2687c4..8d42e2f 100644
--- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h
+++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Burst.h
@@ -46,6 +46,11 @@
                                             bool measureTiming, int64_t deadlineNs,
                                             int64_t loopTimeoutDurationNs,
                                             ExecutionResult* executionResult) override;
+    ndk::ScopedAStatus executeSynchronouslyWithConfig(
+            const Request& request, const std::vector<int64_t>& memoryIdentifierTokens,
+            const ExecutionConfig& config, int64_t deadlineNs,
+            ExecutionResult* executionResult) override;
+
     ndk::ScopedAStatus releaseMemoryResource(int64_t memoryIdentifierToken) override;
 
     class ThreadSafeMemoryCache {
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h
index aa29d63..c94f270 100644
--- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h
+++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/Device.h
@@ -31,6 +31,7 @@
 #include <aidl/android/hardware/neuralnetworks/IPreparedModelParcel.h>
 #include <aidl/android/hardware/neuralnetworks/Model.h>
 #include <aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.h>
+#include <aidl/android/hardware/neuralnetworks/PrepareModelConfig.h>
 #include <aidl/android/hardware/neuralnetworks/Priority.h>
 #include <android/binder_auto_utils.h>
 #include <nnapi/IDevice.h>
@@ -72,6 +73,9 @@
             const std::vector<ndk::ScopedFileDescriptor>& dataCache,
             const std::vector<uint8_t>& token,
             const std::shared_ptr<IPreparedModelCallback>& callback) override;
+    ndk::ScopedAStatus prepareModelWithConfig(
+            const Model& model, const PrepareModelConfig& config,
+            const std::shared_ptr<IPreparedModelCallback>& callback) override;
 
   protected:
     const ::android::nn::SharedDevice kDevice;
diff --git a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
index f92b0bc..d1359d6 100644
--- a/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
+++ b/neuralnetworks/utils/adapter/aidl/include/nnapi/hal/aidl/PreparedModel.h
@@ -51,9 +51,17 @@
                                      int64_t loopTimeoutDurationNs, int64_t durationNs,
                                      FencedExecutionResult* executionResult) override;
     ndk::ScopedAStatus configureExecutionBurst(std::shared_ptr<IBurst>* burst) override;
-    ndk::ScopedAStatus createReusableExecution(const Request& request, bool measureTiming,
-                                               int64_t loopTimeoutDurationNs,
+    ndk::ScopedAStatus createReusableExecution(const Request& request,
+                                               const ExecutionConfig& config,
                                                std::shared_ptr<IExecution>* execution) override;
+    ndk::ScopedAStatus executeSynchronouslyWithConfig(const Request& request,
+                                                      const ExecutionConfig& config,
+                                                      int64_t deadlineNs,
+                                                      ExecutionResult* executionResult) override;
+    ndk::ScopedAStatus executeFencedWithConfig(
+            const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+            const ExecutionConfig& config, int64_t deadlineNs, int64_t durationNs,
+            FencedExecutionResult* executionResult) override;
 
     ::android::nn::SharedPreparedModel getUnderlyingPreparedModel() const;
 
diff --git a/neuralnetworks/utils/adapter/aidl/src/Burst.cpp b/neuralnetworks/utils/adapter/aidl/src/Burst.cpp
index 4fabb20..a4a80fa 100644
--- a/neuralnetworks/utils/adapter/aidl/src/Burst.cpp
+++ b/neuralnetworks/utils/adapter/aidl/src/Burst.cpp
@@ -93,7 +93,8 @@
 nn::ExecutionResult<ExecutionResult> executeSynchronously(
         const nn::IBurst& burst, const Burst::ThreadSafeMemoryCache& cache, const Request& request,
         const std::vector<int64_t>& memoryIdentifierTokens, bool measureTiming, int64_t deadlineNs,
-        int64_t loopTimeoutDurationNs) {
+        int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints,
+        const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) {
     if (request.pools.size() != memoryIdentifierTokens.size()) {
         return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT)
                << "request.pools.size() != memoryIdentifierTokens.size()";
@@ -107,11 +108,13 @@
     const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO;
     const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs));
     const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs));
+    auto nnHints = NN_TRY(convertInput(hints));
+    auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix));
 
     const auto hold = ensureAllMemoriesAreCached(&nnRequest, memoryIdentifierTokens, burst, cache);
 
-    const auto result =
-            burst.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration);
+    const auto result = burst.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration,
+                                      nnHints, nnExtensionNameToPrefix);
 
     if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) {
         const auto& [message, code, outputShapes] = result.error();
@@ -155,7 +158,24 @@
                                                ExecutionResult* executionResult) {
     auto result =
             adapter::executeSynchronously(*kBurst, kMemoryCache, request, memoryIdentifierTokens,
-                                          measureTiming, deadlineNs, loopTimeoutDurationNs);
+                                          measureTiming, deadlineNs, loopTimeoutDurationNs, {}, {});
+    if (!result.has_value()) {
+        auto [message, code, _] = std::move(result).error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *executionResult = std::move(result).value();
+    return ndk::ScopedAStatus::ok();
+}
+
+ndk::ScopedAStatus Burst::executeSynchronouslyWithConfig(
+        const Request& request, const std::vector<int64_t>& memoryIdentifierTokens,
+        const ExecutionConfig& config, int64_t deadlineNs, ExecutionResult* executionResult) {
+    auto result = adapter::executeSynchronously(
+            *kBurst, kMemoryCache, request, memoryIdentifierTokens, config.measureTiming,
+            deadlineNs, config.loopTimeoutDurationNs, config.executionHints,
+            config.extensionNameToPrefix);
     if (!result.has_value()) {
         auto [message, code, _] = std::move(result).error();
         const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
diff --git a/neuralnetworks/utils/adapter/aidl/src/Device.cpp b/neuralnetworks/utils/adapter/aidl/src/Device.cpp
index 763be7f..84aaddb 100644
--- a/neuralnetworks/utils/adapter/aidl/src/Device.cpp
+++ b/neuralnetworks/utils/adapter/aidl/src/Device.cpp
@@ -148,13 +148,14 @@
     }
 }
 
-nn::GeneralResult<void> prepareModel(const nn::SharedDevice& device, const Executor& executor,
-                                     const Model& model, ExecutionPreference preference,
-                                     Priority priority, int64_t deadlineNs,
-                                     const std::vector<ndk::ScopedFileDescriptor>& modelCache,
-                                     const std::vector<ndk::ScopedFileDescriptor>& dataCache,
-                                     const std::vector<uint8_t>& token,
-                                     const std::shared_ptr<IPreparedModelCallback>& callback) {
+nn::GeneralResult<void> prepareModel(
+        const nn::SharedDevice& device, const Executor& executor, const Model& model,
+        ExecutionPreference preference, Priority priority, int64_t deadlineNs,
+        const std::vector<ndk::ScopedFileDescriptor>& modelCache,
+        const std::vector<ndk::ScopedFileDescriptor>& dataCache, const std::vector<uint8_t>& token,
+        const std::vector<TokenValuePair>& hints,
+        const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix,
+        const std::shared_ptr<IPreparedModelCallback>& callback) {
     if (callback.get() == nullptr) {
         return NN_ERROR(nn::ErrorStatus::INVALID_ARGUMENT) << "Invalid callback";
     }
@@ -166,12 +167,16 @@
     auto nnModelCache = NN_TRY(convertInput(modelCache));
     auto nnDataCache = NN_TRY(convertInput(dataCache));
     const auto nnToken = NN_TRY(convertCacheToken(token));
+    auto nnHints = NN_TRY(convertInput(hints));
+    auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix));
 
     Task task = [device, nnModel = std::move(nnModel), nnPreference, nnPriority, nnDeadline,
                  nnModelCache = std::move(nnModelCache), nnDataCache = std::move(nnDataCache),
-                 nnToken, callback] {
-        auto result = device->prepareModel(nnModel, nnPreference, nnPriority, nnDeadline,
-                                           nnModelCache, nnDataCache, nnToken);
+                 nnToken, nnHints = std::move(nnHints),
+                 nnExtensionNameToPrefix = std::move(nnExtensionNameToPrefix), callback] {
+        auto result =
+                device->prepareModel(nnModel, nnPreference, nnPriority, nnDeadline, nnModelCache,
+                                     nnDataCache, nnToken, nnHints, nnExtensionNameToPrefix);
         notify(callback.get(), std::move(result));
     };
     executor(std::move(task), nnDeadline);
@@ -273,8 +278,9 @@
                                         const std::vector<ndk::ScopedFileDescriptor>& dataCache,
                                         const std::vector<uint8_t>& token,
                                         const std::shared_ptr<IPreparedModelCallback>& callback) {
-    const auto result = adapter::prepareModel(kDevice, kExecutor, model, preference, priority,
-                                              deadlineNs, modelCache, dataCache, token, callback);
+    const auto result =
+            adapter::prepareModel(kDevice, kExecutor, model, preference, priority, deadlineNs,
+                                  modelCache, dataCache, token, {}, {}, callback);
     if (!result.has_value()) {
         const auto& [message, code] = result.error();
         const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
@@ -301,4 +307,21 @@
     return ndk::ScopedAStatus::ok();
 }
 
+ndk::ScopedAStatus Device::prepareModelWithConfig(
+        const Model& model, const PrepareModelConfig& config,
+        const std::shared_ptr<IPreparedModelCallback>& callback) {
+    const auto result = adapter::prepareModel(
+            kDevice, kExecutor, model, config.preference, config.priority, config.deadlineNs,
+            config.modelCache, config.dataCache, config.cacheToken, config.compilationHints,
+            config.extensionNameToPrefix, callback);
+    if (!result.has_value()) {
+        const auto& [message, code] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        callback->notify(aidlCode, nullptr);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    return ndk::ScopedAStatus::ok();
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::adapter
diff --git a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
index 5cab62c..790558f 100644
--- a/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
+++ b/neuralnetworks/utils/adapter/aidl/src/PreparedModel.cpp
@@ -118,17 +118,20 @@
     return durationNs < 0 ? nn::OptionalTimePoint{} : nn::TimePoint(makeDuration(durationNs));
 }
 
-nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IPreparedModel& preparedModel,
-                                                          const Request& request,
-                                                          bool measureTiming, int64_t deadlineNs,
-                                                          int64_t loopTimeoutDurationNs) {
+nn::ExecutionResult<ExecutionResult> executeSynchronously(
+        const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming,
+        int64_t deadlineNs, int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints,
+        const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) {
     const auto nnRequest = NN_TRY(convertInput(request));
     const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO;
     const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs));
     const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs));
+    auto nnHints = NN_TRY(convertInput(hints));
+    auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix));
 
     const auto result =
-            preparedModel.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration);
+            preparedModel.execute(nnRequest, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration,
+                                  nnHints, nnExtensionNameToPrefix);
 
     if (!result.ok() && result.error().code == nn::ErrorStatus::OUTPUT_INSUFFICIENT_SIZE) {
         const auto& [message, code, outputShapes] = result.error();
@@ -147,16 +150,21 @@
 nn::GeneralResult<FencedExecutionResult> executeFenced(
         const nn::IPreparedModel& preparedModel, const Request& request,
         const std::vector<ndk::ScopedFileDescriptor>& waitFor, bool measureTiming,
-        int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs) {
+        int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs,
+        const std::vector<TokenValuePair>& hints,
+        const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) {
     const auto nnRequest = NN_TRY(convertInput(request));
     const auto nnWaitFor = NN_TRY(convertSyncFences(waitFor));
     const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO;
     const auto nnDeadline = NN_TRY(makeOptionalTimePoint(deadlineNs));
     const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs));
     const auto nnDuration = NN_TRY(makeOptionalDuration(durationNs));
+    auto nnHints = NN_TRY(convertInput(hints));
+    auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix));
 
     auto [syncFence, executeFencedInfoCallback] = NN_TRY(preparedModel.executeFenced(
-            nnRequest, nnWaitFor, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, nnDuration));
+            nnRequest, nnWaitFor, nnMeasureTiming, nnDeadline, nnLoopTimeoutDuration, nnDuration,
+            nnHints, nnExtensionNameToPrefix));
 
     ndk::ScopedFileDescriptor fileDescriptor;
     if (syncFence.hasFd()) {
@@ -171,11 +179,16 @@
 
 nn::GeneralResult<nn::SharedExecution> createReusableExecution(
         const nn::IPreparedModel& preparedModel, const Request& request, bool measureTiming,
-        int64_t loopTimeoutDurationNs) {
+        int64_t loopTimeoutDurationNs, const std::vector<TokenValuePair>& hints,
+        const std::vector<ExtensionNameAndPrefix>& extensionNameToPrefix) {
     const auto nnRequest = NN_TRY(convertInput(request));
     const auto nnMeasureTiming = measureTiming ? nn::MeasureTiming::YES : nn::MeasureTiming::NO;
     const auto nnLoopTimeoutDuration = NN_TRY(makeOptionalDuration(loopTimeoutDurationNs));
-    return preparedModel.createReusableExecution(nnRequest, nnMeasureTiming, nnLoopTimeoutDuration);
+    auto nnHints = NN_TRY(convertInput(hints));
+    auto nnExtensionNameToPrefix = NN_TRY(convertInput(extensionNameToPrefix));
+
+    return preparedModel.createReusableExecution(nnRequest, nnMeasureTiming, nnLoopTimeoutDuration,
+                                                 nnHints, nnExtensionNameToPrefix);
 }
 
 nn::ExecutionResult<ExecutionResult> executeSynchronously(const nn::IExecution& execution,
@@ -231,7 +244,7 @@
                                                        int64_t loopTimeoutDurationNs,
                                                        ExecutionResult* executionResult) {
     auto result = adapter::executeSynchronously(*kPreparedModel, request, measureTiming, deadlineNs,
-                                                loopTimeoutDurationNs);
+                                                loopTimeoutDurationNs, {}, {});
     if (!result.has_value()) {
         const auto& [message, code, _] = result.error();
         const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
@@ -247,7 +260,41 @@
         bool measureTiming, int64_t deadlineNs, int64_t loopTimeoutDurationNs, int64_t durationNs,
         FencedExecutionResult* executionResult) {
     auto result = adapter::executeFenced(*kPreparedModel, request, waitFor, measureTiming,
-                                         deadlineNs, loopTimeoutDurationNs, durationNs);
+                                         deadlineNs, loopTimeoutDurationNs, durationNs, {}, {});
+    if (!result.has_value()) {
+        const auto& [message, code] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *executionResult = std::move(result).value();
+    return ndk::ScopedAStatus::ok();
+}
+
+ndk::ScopedAStatus PreparedModel::executeSynchronouslyWithConfig(const Request& request,
+                                                                 const ExecutionConfig& config,
+                                                                 int64_t deadlineNs,
+                                                                 ExecutionResult* executionResult) {
+    auto result = adapter::executeSynchronously(
+            *kPreparedModel, request, config.measureTiming, deadlineNs,
+            config.loopTimeoutDurationNs, config.executionHints, config.extensionNameToPrefix);
+    if (!result.has_value()) {
+        const auto& [message, code, _] = result.error();
+        const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
+        return ndk::ScopedAStatus::fromServiceSpecificErrorWithMessage(
+                static_cast<int32_t>(aidlCode), message.c_str());
+    }
+    *executionResult = std::move(result).value();
+    return ndk::ScopedAStatus::ok();
+}
+
+ndk::ScopedAStatus PreparedModel::executeFencedWithConfig(
+        const Request& request, const std::vector<ndk::ScopedFileDescriptor>& waitFor,
+        const ExecutionConfig& config, int64_t deadlineNs, int64_t durationNs,
+        FencedExecutionResult* executionResult) {
+    auto result = adapter::executeFenced(*kPreparedModel, request, waitFor, config.measureTiming,
+                                         deadlineNs, config.loopTimeoutDurationNs, durationNs,
+                                         config.executionHints, config.extensionNameToPrefix);
     if (!result.has_value()) {
         const auto& [message, code] = result.error();
         const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
@@ -275,11 +322,11 @@
 }
 
 ndk::ScopedAStatus PreparedModel::createReusableExecution(const Request& request,
-                                                          bool measureTiming,
-                                                          int64_t loopTimeoutDurationNs,
+                                                          const ExecutionConfig& config,
                                                           std::shared_ptr<IExecution>* execution) {
-    auto result = adapter::createReusableExecution(*kPreparedModel, request, measureTiming,
-                                                   loopTimeoutDurationNs);
+    auto result = adapter::createReusableExecution(
+            *kPreparedModel, request, config.measureTiming, config.loopTimeoutDurationNs,
+            config.executionHints, config.extensionNameToPrefix);
     if (!result.has_value()) {
         const auto& [message, code] = result.error();
         const auto aidlCode = utils::convert(code).value_or(ErrorStatus::GENERAL_FAILURE);
diff --git a/neuralnetworks/utils/adapter/hidl/Android.bp b/neuralnetworks/utils/adapter/hidl/Android.bp
index d073106..6875daa 100644
--- a/neuralnetworks/utils/adapter/hidl/Android.bp
+++ b/neuralnetworks/utils/adapter/hidl/Android.bp
@@ -30,17 +30,16 @@
     local_include_dirs: ["include/nnapi/hal"],
     export_include_dirs: ["include"],
     static_libs: [
-        "neuralnetworks_types",
-        "neuralnetworks_utils_hal_1_0",
-        "neuralnetworks_utils_hal_1_1",
-        "neuralnetworks_utils_hal_1_2",
-        "neuralnetworks_utils_hal_1_3",
-    ],
-    shared_libs: [
         "android.hardware.neuralnetworks@1.0",
         "android.hardware.neuralnetworks@1.1",
         "android.hardware.neuralnetworks@1.2",
         "android.hardware.neuralnetworks@1.3",
         "libfmq",
+        "neuralnetworks_types",
+        "neuralnetworks_utils_hal_1_0",
+        "neuralnetworks_utils_hal_1_1",
+        "neuralnetworks_utils_hal_1_2",
+        "neuralnetworks_utils_hal_1_3",
+        "neuralnetworks_utils_hal_common",
     ],
 }
diff --git a/neuralnetworks/utils/adapter/hidl/src/Burst.cpp b/neuralnetworks/utils/adapter/hidl/src/Burst.cpp
index 8b2e1dd..e3b165b 100644
--- a/neuralnetworks/utils/adapter/hidl/src/Burst.cpp
+++ b/neuralnetworks/utils/adapter/hidl/src/Burst.cpp
@@ -250,7 +250,7 @@
     nn::MeasureTiming canonicalMeasure = NN_TRY(nn::convert(measure));
 
     const auto [outputShapes, timing] =
-            NN_TRY(mBurstExecutor->execute(canonicalRequest, canonicalMeasure, {}, {}));
+            NN_TRY(mBurstExecutor->execute(canonicalRequest, canonicalMeasure, {}, {}, {}, {}));
 
     return std::make_pair(NN_TRY(V1_2::utils::convert(outputShapes)),
                           NN_TRY(V1_2::utils::convert(timing)));
diff --git a/neuralnetworks/utils/adapter/hidl/src/Device.cpp b/neuralnetworks/utils/adapter/hidl/src/Device.cpp
index 4993a80..0f44638 100644
--- a/neuralnetworks/utils/adapter/hidl/src/Device.cpp
+++ b/neuralnetworks/utils/adapter/hidl/src/Device.cpp
@@ -135,7 +135,7 @@
 
     Task task = [device, nnModel = std::move(nnModel), executor, callback] {
         auto result = device->prepareModel(nnModel, nn::ExecutionPreference::DEFAULT,
-                                           nn::Priority::DEFAULT, {}, {}, {}, {});
+                                           nn::Priority::DEFAULT, {}, {}, {}, {}, {}, {});
         notify(callback.get(), std::move(result), executor);
     };
     executor(std::move(task), {});
@@ -155,8 +155,8 @@
     const auto nnPreference = NN_TRY(convertInput(preference));
 
     Task task = [device, nnModel = std::move(nnModel), nnPreference, executor, callback] {
-        auto result =
-                device->prepareModel(nnModel, nnPreference, nn::Priority::DEFAULT, {}, {}, {}, {});
+        auto result = device->prepareModel(nnModel, nnPreference, nn::Priority::DEFAULT, {}, {}, {},
+                                           {}, {}, {});
         notify(callback.get(), std::move(result), executor);
     };
     executor(std::move(task), {});
@@ -185,7 +185,7 @@
                  nnModelCache = std::move(nnModelCache), nnDataCache = std::move(nnDataCache),
                  nnToken, executor, callback] {
         auto result = device->prepareModel(nnModel, nnPreference, nn::Priority::DEFAULT, {},
-                                           nnModelCache, nnDataCache, nnToken);
+                                           nnModelCache, nnDataCache, nnToken, {}, {});
         notify(callback.get(), std::move(result), executor);
     };
     executor(std::move(task), {});
@@ -215,7 +215,7 @@
                  nnModelCache = std::move(nnModelCache), nnDataCache = std::move(nnDataCache),
                  nnToken, executor, callback] {
         auto result = device->prepareModel(nnModel, nnPreference, nnPriority, nnDeadline,
-                                           nnModelCache, nnDataCache, nnToken);
+                                           nnModelCache, nnDataCache, nnToken, {}, {});
         notify(callback.get(), std::move(result), executor);
     };
     executor(std::move(task), nnDeadline);
diff --git a/neuralnetworks/utils/adapter/hidl/src/PreparedModel.cpp b/neuralnetworks/utils/adapter/hidl/src/PreparedModel.cpp
index 71060d5..c6055a6 100644
--- a/neuralnetworks/utils/adapter/hidl/src/PreparedModel.cpp
+++ b/neuralnetworks/utils/adapter/hidl/src/PreparedModel.cpp
@@ -159,7 +159,7 @@
     }
 
     Task task = [preparedModel, nnRequest = std::move(nnRequest), callback] {
-        auto result = preparedModel->execute(nnRequest, nn::MeasureTiming::NO, {}, {});
+        auto result = preparedModel->execute(nnRequest, nn::MeasureTiming::NO, {}, {}, {}, {});
         notify(callback.get(), std::move(result));
     };
     executor(std::move(task), {});
@@ -185,7 +185,7 @@
     }
 
     Task task = [preparedModel, nnRequest = std::move(nnRequest), nnMeasure, callback] {
-        auto result = preparedModel->execute(nnRequest, nnMeasure, {}, {});
+        auto result = preparedModel->execute(nnRequest, nnMeasure, {}, {}, {}, {});
         notify(callback.get(), std::move(result));
     };
     executor(std::move(task), {});
@@ -216,8 +216,8 @@
 
     Task task = [preparedModel, nnRequest = std::move(nnRequest), nnMeasure, nnDeadline,
                  nnLoopTimeoutDuration, callback] {
-        auto result =
-                preparedModel->execute(nnRequest, nnMeasure, nnDeadline, nnLoopTimeoutDuration);
+        auto result = preparedModel->execute(nnRequest, nnMeasure, nnDeadline,
+                                             nnLoopTimeoutDuration, {}, {});
         notify(callback.get(), std::move(result));
     };
     executor(std::move(task), nnDeadline);
@@ -232,7 +232,7 @@
     const auto nnMeasure = NN_TRY(convertInput(measure));
 
     const auto [outputShapes, timing] =
-            NN_TRY(preparedModel->execute(nnRequest, nnMeasure, {}, {}));
+            NN_TRY(preparedModel->execute(nnRequest, nnMeasure, {}, {}, {}, {}));
 
     auto hidlOutputShapes = NN_TRY(V1_2::utils::convert(outputShapes));
     const auto hidlTiming = NN_TRY(V1_2::utils::convert(timing));
@@ -248,8 +248,8 @@
     const auto nnDeadline = NN_TRY(convertInput(deadline));
     const auto nnLoopTimeoutDuration = NN_TRY(convertInput(loopTimeoutDuration));
 
-    const auto [outputShapes, timing] =
-            NN_TRY(preparedModel->execute(nnRequest, nnMeasure, nnDeadline, nnLoopTimeoutDuration));
+    const auto [outputShapes, timing] = NN_TRY(preparedModel->execute(
+            nnRequest, nnMeasure, nnDeadline, nnLoopTimeoutDuration, {}, {}));
 
     auto hidlOutputShapes = NN_TRY(V1_3::utils::convert(outputShapes));
     const auto hidlTiming = NN_TRY(V1_3::utils::convert(timing));
@@ -293,8 +293,9 @@
     const auto nnLoopTimeoutDuration = NN_TRY(convertInput(loopTimeoutDuration));
     const auto nnDuration = NN_TRY(convertInput(duration));
 
-    auto [syncFence, executeFencedCallback] = NN_TRY(preparedModel->executeFenced(
-            nnRequest, nnWaitFor, nnMeasure, nnDeadline, nnLoopTimeoutDuration, nnDuration));
+    auto [syncFence, executeFencedCallback] =
+            NN_TRY(preparedModel->executeFenced(nnRequest, nnWaitFor, nnMeasure, nnDeadline,
+                                                nnLoopTimeoutDuration, nnDuration, {}, {}));
 
     auto hidlSyncFence = NN_TRY(V1_3::utils::convert(syncFence.getSharedHandle()));
     auto hidlExecuteFencedCallback = sp<FencedExecutionCallback>::make(executeFencedCallback);
diff --git a/neuralnetworks/utils/common/Android.bp b/neuralnetworks/utils/common/Android.bp
index 39927a3..bfba24f 100644
--- a/neuralnetworks/utils/common/Android.bp
+++ b/neuralnetworks/utils/common/Android.bp
@@ -39,20 +39,12 @@
     srcs: ["test/*.cpp"],
     static_libs: [
         "libgmock",
-        "libneuralnetworks_common",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_common",
     ],
     shared_libs: [
-        "android.hidl.allocator@1.0",
-        "android.hidl.memory@1.0",
         "libbase",
         "libcutils",
-        "libfmq",
-        "libhidlbase",
-        "libhidlmemory",
-        "liblog",
-        "libutils",
     ],
     target: {
         android: {
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/InvalidBurst.h b/neuralnetworks/utils/common/include/nnapi/hal/InvalidBurst.h
index e86edda..1f1245f 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/InvalidBurst.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/InvalidBurst.h
@@ -33,12 +33,15 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 };
 
 }  // namespace android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/InvalidDevice.h b/neuralnetworks/utils/common/include/nnapi/hal/InvalidDevice.h
index 5e62b9a..9582873 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/InvalidDevice.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/InvalidDevice.h
@@ -52,8 +52,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/InvalidPreparedModel.h b/neuralnetworks/utils/common/include/nnapi/hal/InvalidPreparedModel.h
index de30aae..3f1f290 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/InvalidPreparedModel.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/InvalidPreparedModel.h
@@ -31,18 +31,23 @@
   public:
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/ResilientBurst.h b/neuralnetworks/utils/common/include/nnapi/hal/ResilientBurst.h
index fde2486..129431f 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/ResilientBurst.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/ResilientBurst.h
@@ -48,18 +48,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
   private:
     bool isValidInternal() const EXCLUDES(mMutex);
     nn::GeneralResult<nn::SharedExecution> createReusableExecutionInternal(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const;
 
     const Factory kMakeBurst;
     mutable std::mutex mMutex;
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/ResilientDevice.h b/neuralnetworks/utils/common/include/nnapi/hal/ResilientDevice.h
index 84ae799..267d634 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/ResilientDevice.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/ResilientDevice.h
@@ -65,8 +65,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModel(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache,
-            const nn::CacheToken& token) const override;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCache(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
@@ -83,7 +84,9 @@
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelInternal(
             const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const;
+            const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const;
     nn::GeneralResult<nn::SharedPreparedModel> prepareModelFromCacheInternal(
             nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
             const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const;
diff --git a/neuralnetworks/utils/common/include/nnapi/hal/ResilientPreparedModel.h b/neuralnetworks/utils/common/include/nnapi/hal/ResilientPreparedModel.h
index 86533ed..bbfc220 100644
--- a/neuralnetworks/utils/common/include/nnapi/hal/ResilientPreparedModel.h
+++ b/neuralnetworks/utils/common/include/nnapi/hal/ResilientPreparedModel.h
@@ -49,18 +49,23 @@
 
     nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> execute(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalTimePoint& deadline,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>> executeFenced(
             const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
             nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
             const nn::OptionalDuration& loopTimeoutDuration,
-            const nn::OptionalDuration& timeoutDurationAfterFence) const override;
+            const nn::OptionalDuration& timeoutDurationAfterFence,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedExecution> createReusableExecution(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const override;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& hints,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const override;
 
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurst() const override;
 
@@ -70,7 +75,9 @@
     bool isValidInternal() const EXCLUDES(mMutex);
     nn::GeneralResult<nn::SharedExecution> createReusableExecutionInternal(
             const nn::Request& request, nn::MeasureTiming measure,
-            const nn::OptionalDuration& loopTimeoutDuration) const;
+            const nn::OptionalDuration& loopTimeoutDuration,
+            const std::vector<nn::TokenValuePair>& metaData,
+            const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const;
     nn::GeneralResult<nn::SharedBurst> configureExecutionBurstInternal() const;
 
     const Factory kMakePreparedModel;
diff --git a/neuralnetworks/utils/common/src/InvalidBurst.cpp b/neuralnetworks/utils/common/src/InvalidBurst.cpp
index 0191533..3fdfb5c 100644
--- a/neuralnetworks/utils/common/src/InvalidBurst.cpp
+++ b/neuralnetworks/utils/common/src/InvalidBurst.cpp
@@ -34,13 +34,17 @@
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> InvalidBurst::execute(
         const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
         const nn::OptionalTimePoint& /*deadline*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidBurst";
 }
 
 nn::GeneralResult<nn::SharedExecution> InvalidBurst::createReusableExecution(
         const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidBurst";
 }
 
diff --git a/neuralnetworks/utils/common/src/InvalidDevice.cpp b/neuralnetworks/utils/common/src/InvalidDevice.cpp
index 535ccb4..c8cc287 100644
--- a/neuralnetworks/utils/common/src/InvalidDevice.cpp
+++ b/neuralnetworks/utils/common/src/InvalidDevice.cpp
@@ -84,7 +84,9 @@
         const nn::Model& /*model*/, nn::ExecutionPreference /*preference*/,
         nn::Priority /*priority*/, nn::OptionalTimePoint /*deadline*/,
         const std::vector<nn::SharedHandle>& /*modelCache*/,
-        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/) const {
+        const std::vector<nn::SharedHandle>& /*dataCache*/, const nn::CacheToken& /*token*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidDevice";
 }
 
diff --git a/neuralnetworks/utils/common/src/InvalidPreparedModel.cpp b/neuralnetworks/utils/common/src/InvalidPreparedModel.cpp
index 8195462..f6f978d 100644
--- a/neuralnetworks/utils/common/src/InvalidPreparedModel.cpp
+++ b/neuralnetworks/utils/common/src/InvalidPreparedModel.cpp
@@ -27,9 +27,12 @@
 namespace android::hardware::neuralnetworks::utils {
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
-InvalidPreparedModel::execute(const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
-                              const nn::OptionalTimePoint& /*deadline*/,
-                              const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+InvalidPreparedModel::execute(
+        const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
+        const nn::OptionalTimePoint& /*deadline*/,
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidPreparedModel";
 }
 
@@ -38,13 +41,17 @@
         const nn::Request& /*request*/, const std::vector<nn::SyncFence>& /*waitFor*/,
         nn::MeasureTiming /*measure*/, const nn::OptionalTimePoint& /*deadline*/,
         const nn::OptionalDuration& /*loopTimeoutDuration*/,
-        const nn::OptionalDuration& /*timeoutDurationAfterFence*/) const {
+        const nn::OptionalDuration& /*timeoutDurationAfterFence*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidPreparedModel";
 }
 
 nn::GeneralResult<nn::SharedExecution> InvalidPreparedModel::createReusableExecution(
         const nn::Request& /*request*/, nn::MeasureTiming /*measure*/,
-        const nn::OptionalDuration& /*loopTimeoutDuration*/) const {
+        const nn::OptionalDuration& /*loopTimeoutDuration*/,
+        const std::vector<nn::TokenValuePair>& /*hints*/,
+        const std::vector<nn::ExtensionNameAndPrefix>& /*extensionNameToPrefix*/) const {
     return NN_ERROR() << "InvalidPreparedModel";
 }
 
diff --git a/neuralnetworks/utils/common/src/ResilientBurst.cpp b/neuralnetworks/utils/common/src/ResilientBurst.cpp
index 79cbe39..bf7a8ea 100644
--- a/neuralnetworks/utils/common/src/ResilientBurst.cpp
+++ b/neuralnetworks/utils/common/src/ResilientBurst.cpp
@@ -105,37 +105,49 @@
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>> ResilientBurst::execute(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalTimePoint& deadline,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
-    const auto fn = [&request, measure, deadline, loopTimeoutDuration](const nn::IBurst& burst) {
-        return burst.execute(request, measure, deadline, loopTimeoutDuration);
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
+    const auto fn = [&request, measure, deadline, loopTimeoutDuration, &hints,
+                     &extensionNameToPrefix](const nn::IBurst& burst) {
+        return burst.execute(request, measure, deadline, loopTimeoutDuration, hints,
+                             extensionNameToPrefix);
     };
     return protect(*this, fn);
 }
 
 nn::GeneralResult<nn::SharedExecution> ResilientBurst::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
 #if 0
     auto self = shared_from_this();
-    ResilientExecution::Factory makeExecution =
-            [burst = std::move(self), request, measure, loopTimeoutDuration] {
-        return burst->createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+    ResilientExecution::Factory makeExecution = [burst = std::move(self), request, measure,
+                                                 loopTimeoutDuration, &hints,
+                                                 &extensionNameToPrefix] {
+        return burst->createReusableExecutionInternal(request, measure, loopTimeoutDuration, hints,
+                                                      extensionNameToPrefix);
     };
     return ResilientExecution::create(std::move(makeExecution));
 #else
-    return createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+    return createReusableExecutionInternal(request, measure, loopTimeoutDuration, hints,
+                                           extensionNameToPrefix);
 #endif
 }
 
 nn::GeneralResult<nn::SharedExecution> ResilientBurst::createReusableExecutionInternal(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     if (!isValidInternal()) {
         return std::make_shared<const InvalidExecution>();
     }
-    const auto fn = [&request, measure, &loopTimeoutDuration](const nn::IBurst& burst) {
-        return burst.createReusableExecution(request, measure, loopTimeoutDuration);
+    const auto fn = [&request, measure, &loopTimeoutDuration, &hints,
+                     &extensionNameToPrefix](const nn::IBurst& burst) {
+        return burst.createReusableExecution(request, measure, loopTimeoutDuration, hints,
+                                             extensionNameToPrefix);
     };
     return protect(*this, fn);
 }
diff --git a/neuralnetworks/utils/common/src/ResilientDevice.cpp b/neuralnetworks/utils/common/src/ResilientDevice.cpp
index 2023c9a..a5c2640 100644
--- a/neuralnetworks/utils/common/src/ResilientDevice.cpp
+++ b/neuralnetworks/utils/common/src/ResilientDevice.cpp
@@ -179,19 +179,21 @@
 nn::GeneralResult<nn::SharedPreparedModel> ResilientDevice::prepareModel(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
         nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
 #if 0
     auto self = shared_from_this();
     ResilientPreparedModel::Factory makePreparedModel = [device = std::move(self), model,
                                                          preference, priority, deadline, modelCache,
-                                                         dataCache, token] {
+                                                         dataCache, token, hints, extensionNameToPrefix] {
         return device->prepareModelInternal(model, preference, priority, deadline, modelCache,
-                                            dataCache, token);
+                                            dataCache, token, hints, extensionNameToPrefix);
     };
     return ResilientPreparedModel::create(std::move(makePreparedModel));
 #else
-    return prepareModelInternal(model, preference, priority, deadline, modelCache, dataCache,
-                                token);
+    return prepareModelInternal(model, preference, priority, deadline, modelCache, dataCache, token,
+                                hints, extensionNameToPrefix);
 #endif
 }
 
@@ -234,14 +236,16 @@
 nn::GeneralResult<nn::SharedPreparedModel> ResilientDevice::prepareModelInternal(
         const nn::Model& model, nn::ExecutionPreference preference, nn::Priority priority,
         nn::OptionalTimePoint deadline, const std::vector<nn::SharedHandle>& modelCache,
-        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token) const {
+        const std::vector<nn::SharedHandle>& dataCache, const nn::CacheToken& token,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     if (!isValidInternal()) {
         return std::make_shared<const InvalidPreparedModel>();
     }
-    const auto fn = [&model, preference, priority, &deadline, &modelCache, &dataCache,
-                     &token](const nn::IDevice& device) {
+    const auto fn = [&model, preference, priority, &deadline, &modelCache, &dataCache, &token,
+                     &hints, &extensionNameToPrefix](const nn::IDevice& device) {
         return device.prepareModel(model, preference, priority, deadline, modelCache, dataCache,
-                                   token);
+                                   token, hints, extensionNameToPrefix);
     };
     return protect(*this, fn, /*blocking=*/false);
 }
diff --git a/neuralnetworks/utils/common/src/ResilientPreparedModel.cpp b/neuralnetworks/utils/common/src/ResilientPreparedModel.cpp
index 1ae19bc..b5843c0 100644
--- a/neuralnetworks/utils/common/src/ResilientPreparedModel.cpp
+++ b/neuralnetworks/utils/common/src/ResilientPreparedModel.cpp
@@ -104,43 +104,53 @@
 }
 
 nn::ExecutionResult<std::pair<std::vector<nn::OutputShape>, nn::Timing>>
-ResilientPreparedModel::execute(const nn::Request& request, nn::MeasureTiming measure,
-                                const nn::OptionalTimePoint& deadline,
-                                const nn::OptionalDuration& loopTimeoutDuration) const {
-    const auto fn = [&request, measure, &deadline,
-                     &loopTimeoutDuration](const nn::IPreparedModel& preparedModel) {
-        return preparedModel.execute(request, measure, deadline, loopTimeoutDuration);
+ResilientPreparedModel::execute(
+        const nn::Request& request, nn::MeasureTiming measure,
+        const nn::OptionalTimePoint& deadline, const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
+    const auto fn = [&request, measure, &deadline, &loopTimeoutDuration, &hints,
+                     &extensionNameToPrefix](const nn::IPreparedModel& preparedModel) {
+        return preparedModel.execute(request, measure, deadline, loopTimeoutDuration, hints,
+                                     extensionNameToPrefix);
     };
     return protect(*this, fn);
 }
 
 nn::GeneralResult<std::pair<nn::SyncFence, nn::ExecuteFencedInfoCallback>>
-ResilientPreparedModel::executeFenced(const nn::Request& request,
-                                      const std::vector<nn::SyncFence>& waitFor,
-                                      nn::MeasureTiming measure,
-                                      const nn::OptionalTimePoint& deadline,
-                                      const nn::OptionalDuration& loopTimeoutDuration,
-                                      const nn::OptionalDuration& timeoutDurationAfterFence) const {
+ResilientPreparedModel::executeFenced(
+        const nn::Request& request, const std::vector<nn::SyncFence>& waitFor,
+        nn::MeasureTiming measure, const nn::OptionalTimePoint& deadline,
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const nn::OptionalDuration& timeoutDurationAfterFence,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     const auto fn = [&request, &waitFor, measure, &deadline, &loopTimeoutDuration,
-                     &timeoutDurationAfterFence](const nn::IPreparedModel& preparedModel) {
+                     &timeoutDurationAfterFence, &hints,
+                     &extensionNameToPrefix](const nn::IPreparedModel& preparedModel) {
         return preparedModel.executeFenced(request, waitFor, measure, deadline, loopTimeoutDuration,
-                                           timeoutDurationAfterFence);
+                                           timeoutDurationAfterFence, hints, extensionNameToPrefix);
     };
     return protect(*this, fn);
 }
 
 nn::GeneralResult<nn::SharedExecution> ResilientPreparedModel::createReusableExecution(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
 #if 0
     auto self = shared_from_this();
-    ResilientExecution::Factory makeExecution =
-            [preparedModel = std::move(self), request, measure, loopTimeoutDuration] {
-        return preparedModel->createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+    ResilientExecution::Factory makeExecution = [preparedModel = std::move(self), request, measure,
+                                                 loopTimeoutDuration, hints,
+                                                 extensionNameToPrefix] {
+        return preparedModel->createReusableExecutionInternal(request, measure, loopTimeoutDuration,
+                                                              hints, extensionNameToPrefix);
     };
     return ResilientExecution::create(std::move(makeExecution));
 #else
-    return createReusableExecutionInternal(request, measure, loopTimeoutDuration);
+    return createReusableExecutionInternal(request, measure, loopTimeoutDuration, hints,
+                                           extensionNameToPrefix);
 #endif
 }
 
@@ -159,13 +169,16 @@
 
 nn::GeneralResult<nn::SharedExecution> ResilientPreparedModel::createReusableExecutionInternal(
         const nn::Request& request, nn::MeasureTiming measure,
-        const nn::OptionalDuration& loopTimeoutDuration) const {
+        const nn::OptionalDuration& loopTimeoutDuration,
+        const std::vector<nn::TokenValuePair>& hints,
+        const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix) const {
     if (!isValidInternal()) {
         return std::make_shared<const InvalidExecution>();
     }
-    const auto fn = [&request, measure,
-                     &loopTimeoutDuration](const nn::IPreparedModel& preparedModel) {
-        return preparedModel.createReusableExecution(request, measure, loopTimeoutDuration);
+    const auto fn = [&request, measure, &loopTimeoutDuration, &hints,
+                     &extensionNameToPrefix](const nn::IPreparedModel& preparedModel) {
+        return preparedModel.createReusableExecution(request, measure, loopTimeoutDuration, hints,
+                                                     extensionNameToPrefix);
     };
     return protect(*this, fn);
 }
diff --git a/neuralnetworks/utils/common/test/MockDevice.h b/neuralnetworks/utils/common/test/MockDevice.h
index a9428bc..a0fc5c3 100644
--- a/neuralnetworks/utils/common/test/MockDevice.h
+++ b/neuralnetworks/utils/common/test/MockDevice.h
@@ -39,7 +39,9 @@
     MOCK_METHOD(GeneralResult<SharedPreparedModel>, prepareModel,
                 (const Model& model, ExecutionPreference preference, Priority priority,
                  OptionalTimePoint deadline, const std::vector<SharedHandle>& modelCache,
-                 const std::vector<SharedHandle>& dataCache, const CacheToken& token),
+                 const std::vector<SharedHandle>& dataCache, const CacheToken& token,
+                 const std::vector<TokenValuePair>& hints,
+                 const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix),
                 (const, override));
     MOCK_METHOD(GeneralResult<SharedPreparedModel>, prepareModelFromCache,
                 (OptionalTimePoint deadline, const std::vector<SharedHandle>& modelCache,
diff --git a/neuralnetworks/utils/common/test/MockPreparedModel.h b/neuralnetworks/utils/common/test/MockPreparedModel.h
index c8ce006..b8613b2 100644
--- a/neuralnetworks/utils/common/test/MockPreparedModel.h
+++ b/neuralnetworks/utils/common/test/MockPreparedModel.h
@@ -27,17 +27,23 @@
   public:
     MOCK_METHOD((ExecutionResult<std::pair<std::vector<OutputShape>, Timing>>), execute,
                 (const Request& request, MeasureTiming measure, const OptionalTimePoint& deadline,
-                 const OptionalDuration& loopTimeoutDuration),
+                 const OptionalDuration& loopTimeoutDuration,
+                 const std::vector<TokenValuePair>& hints,
+                 const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix),
                 (const, override));
     MOCK_METHOD((GeneralResult<std::pair<SyncFence, ExecuteFencedInfoCallback>>), executeFenced,
                 (const Request& request, const std::vector<SyncFence>& waitFor,
                  MeasureTiming measure, const OptionalTimePoint& deadline,
                  const OptionalDuration& loopTimeoutDuration,
-                 const OptionalDuration& timeoutDurationAfterFence),
+                 const OptionalDuration& timeoutDurationAfterFence,
+                 const std::vector<TokenValuePair>& hints,
+                 const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix),
                 (const, override));
     MOCK_METHOD((GeneralResult<SharedExecution>), createReusableExecution,
-                (const nn::Request& request, nn::MeasureTiming measure,
-                 const nn::OptionalDuration& loopTimeoutDuration),
+                (const Request& request, MeasureTiming measure,
+                 const OptionalDuration& loopTimeoutDuration,
+                 const std::vector<TokenValuePair>& hints,
+                 const std::vector<nn::ExtensionNameAndPrefix>& extensionNameToPrefix),
                 (const, override));
     MOCK_METHOD(GeneralResult<SharedBurst>, configureExecutionBurst, (), (const, override));
     MOCK_METHOD(std::any, getUnderlyingResource, (), (const, override));
diff --git a/neuralnetworks/utils/common/test/ResilientDeviceTest.cpp b/neuralnetworks/utils/common/test/ResilientDeviceTest.cpp
index 0488b63..d9b8505 100644
--- a/neuralnetworks/utils/common/test/ResilientDeviceTest.cpp
+++ b/neuralnetworks/utils/common/test/ResilientDeviceTest.cpp
@@ -309,12 +309,12 @@
     // setup call
     const auto [mockDevice, mockDeviceFactory, device] = setup();
     const auto mockPreparedModel = std::make_shared<const nn::MockPreparedModel>();
-    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _))
+    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(mockPreparedModel));
 
     // run test
-    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {});
+    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -324,12 +324,12 @@
 TEST(ResilientDeviceTest, prepareModelError) {
     // setup call
     const auto [mockDevice, mockDeviceFactory, device] = setup();
-    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _))
+    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {});
+    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -339,13 +339,13 @@
 TEST(ResilientDeviceTest, prepareModelDeadObjectFailedRecovery) {
     // setup call
     const auto [mockDevice, mockDeviceFactory, device] = setup();
-    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _))
+    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnDeadObject);
     EXPECT_CALL(*mockDeviceFactory, Call(false)).Times(1).WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {});
+    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -355,18 +355,18 @@
 TEST(ResilientDeviceTest, prepareModelDeadObjectSuccessfulRecovery) {
     // setup call
     const auto [mockDevice, mockDeviceFactory, device] = setup();
-    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _))
+    EXPECT_CALL(*mockDevice, prepareModel(_, _, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnDeadObject);
     const auto recoveredMockDevice = createConfiguredMockDevice();
     const auto mockPreparedModel = std::make_shared<const nn::MockPreparedModel>();
-    EXPECT_CALL(*recoveredMockDevice, prepareModel(_, _, _, _, _, _, _))
+    EXPECT_CALL(*recoveredMockDevice, prepareModel(_, _, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(mockPreparedModel));
     EXPECT_CALL(*mockDeviceFactory, Call(false)).Times(1).WillOnce(Return(recoveredMockDevice));
 
     // run test
-    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {});
+    const auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -679,7 +679,7 @@
     device->recover(mockDevice.get(), /*blocking=*/false);
 
     // run test
-    auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {});
+    auto result = device->prepareModel({}, {}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
diff --git a/neuralnetworks/utils/common/test/ResilientPreparedModelTest.cpp b/neuralnetworks/utils/common/test/ResilientPreparedModelTest.cpp
index d396ca8..276bfba 100644
--- a/neuralnetworks/utils/common/test/ResilientPreparedModelTest.cpp
+++ b/neuralnetworks/utils/common/test/ResilientPreparedModelTest.cpp
@@ -104,12 +104,12 @@
 TEST(ResilientPreparedModelTest, execute) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(kNoExecutionError));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -119,10 +119,12 @@
 TEST(ResilientPreparedModelTest, executeError) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _)).Times(1).WillOnce(kReturnGeneralFailure);
+    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _, _, _))
+            .Times(1)
+            .WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -132,12 +134,12 @@
 TEST(ResilientPreparedModelTest, executeDeadObjectFailedRecovery) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _)).Times(1).WillOnce(kReturnDeadObject);
+    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _, _, _)).Times(1).WillOnce(kReturnDeadObject);
     constexpr auto ret = [] { return nn::error(nn::ErrorStatus::GENERAL_FAILURE); };
     EXPECT_CALL(*mockPreparedModelFactory, Call()).Times(1).WillOnce(ret);
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -147,9 +149,9 @@
 TEST(ResilientPreparedModelTest, executeDeadObjectSuccessfulRecovery) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _)).Times(1).WillOnce(kReturnDeadObject);
+    EXPECT_CALL(*mockPreparedModel, execute(_, _, _, _, _, _)).Times(1).WillOnce(kReturnDeadObject);
     const auto recoveredMockPreparedModel = createConfiguredMockPreparedModel();
-    EXPECT_CALL(*recoveredMockPreparedModel, execute(_, _, _, _))
+    EXPECT_CALL(*recoveredMockPreparedModel, execute(_, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(kNoExecutionError));
     EXPECT_CALL(*mockPreparedModelFactory, Call())
@@ -157,7 +159,7 @@
             .WillOnce(Return(recoveredMockPreparedModel));
 
     // run test
-    const auto result = preparedModel->execute({}, {}, {}, {});
+    const auto result = preparedModel->execute({}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -167,12 +169,12 @@
 TEST(ResilientPreparedModelTest, executeFenced) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(kNoFencedExecutionError));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -182,12 +184,12 @@
 TEST(ResilientPreparedModelTest, executeFencedError) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -197,13 +199,13 @@
 TEST(ResilientPreparedModelTest, executeFencedDeadObjectFailedRecovery) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnDeadObject);
     EXPECT_CALL(*mockPreparedModelFactory, Call()).Times(1).WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
@@ -213,11 +215,11 @@
 TEST(ResilientPreparedModelTest, executeFencedDeadObjectSuccessfulRecovery) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _))
+    EXPECT_CALL(*mockPreparedModel, executeFenced(_, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnDeadObject);
     const auto recoveredMockPreparedModel = createConfiguredMockPreparedModel();
-    EXPECT_CALL(*recoveredMockPreparedModel, executeFenced(_, _, _, _, _, _))
+    EXPECT_CALL(*recoveredMockPreparedModel, executeFenced(_, _, _, _, _, _, _, _))
             .Times(1)
             .WillOnce(Return(kNoFencedExecutionError));
     EXPECT_CALL(*mockPreparedModelFactory, Call())
@@ -225,7 +227,7 @@
             .WillOnce(Return(recoveredMockPreparedModel));
 
     // run test
-    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {});
+    const auto result = preparedModel->executeFenced({}, {}, {}, {}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -235,12 +237,12 @@
 TEST(ResilientPreparedModelTest, createReusableExecution) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _, _))
             .Times(1)
             .WillOnce(Return(kNoCreateReusableExecutionError));
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_TRUE(result.has_value())
@@ -250,12 +252,12 @@
 TEST(ResilientPreparedModelTest, createReusableExecutionError) {
     // setup call
     const auto [mockPreparedModel, mockPreparedModelFactory, preparedModel] = setup();
-    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _))
+    EXPECT_CALL(*mockPreparedModel, createReusableExecution(_, _, _, _, _))
             .Times(1)
             .WillOnce(kReturnGeneralFailure);
 
     // run test
-    const auto result = preparedModel->createReusableExecution({}, {}, {});
+    const auto result = preparedModel->createReusableExecution({}, {}, {}, {}, {});
 
     // verify result
     ASSERT_FALSE(result.has_value());
diff --git a/neuralnetworks/utils/service/Android.bp b/neuralnetworks/utils/service/Android.bp
index c3272ae..452078b 100644
--- a/neuralnetworks/utils/service/Android.bp
+++ b/neuralnetworks/utils/service/Android.bp
@@ -33,6 +33,10 @@
     local_include_dirs: ["include/nnapi/hal"],
     export_include_dirs: ["include"],
     static_libs: [
+        "android.hardware.neuralnetworks@1.0",
+        "android.hardware.neuralnetworks@1.1",
+        "android.hardware.neuralnetworks@1.2",
+        "android.hardware.neuralnetworks@1.3",
         "neuralnetworks_types",
         "neuralnetworks_utils_hal_1_0",
         "neuralnetworks_utils_hal_1_1",
@@ -40,10 +44,4 @@
         "neuralnetworks_utils_hal_1_3",
         "neuralnetworks_utils_hal_common",
     ],
-    shared_libs: [
-        "android.hardware.neuralnetworks@1.0",
-        "android.hardware.neuralnetworks@1.1",
-        "android.hardware.neuralnetworks@1.2",
-        "android.hardware.neuralnetworks@1.3",
-    ],
 }