diff --git a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
index af58715..71a28ef 100644
--- a/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
+++ b/neuralnetworks/aidl/utils/include/nnapi/hal/aidl/Conversions.h
@@ -27,6 +27,7 @@
 #include <aidl/android/hardware/neuralnetworks/Extension.h>
 #include <aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.h>
 #include <aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.h>
+#include <aidl/android/hardware/neuralnetworks/IDevice.h>
 #include <aidl/android/hardware/neuralnetworks/Memory.h>
 #include <aidl/android/hardware/neuralnetworks/Model.h>
 #include <aidl/android/hardware/neuralnetworks/Operand.h>
@@ -219,6 +220,7 @@
 #endif  // NN_AIDL_V4_OR_ABOVE
 
 nn::GeneralResult<std::vector<int32_t>> toSigned(const std::vector<uint32_t>& vec);
+std::vector<uint8_t> toVec(const std::array<uint8_t, IDevice::BYTE_SIZE_OF_CACHE_TOKEN>& token);
 
 }  // namespace aidl::android::hardware::neuralnetworks::utils
 
diff --git a/neuralnetworks/aidl/utils/src/Conversions.cpp b/neuralnetworks/aidl/utils/src/Conversions.cpp
index 9b897c4..83fda10 100644
--- a/neuralnetworks/aidl/utils/src/Conversions.cpp
+++ b/neuralnetworks/aidl/utils/src/Conversions.cpp
@@ -614,7 +614,7 @@
     using Ts::operator()...;
 };
 template <class... Ts>
-overloaded(Ts...)->overloaded<Ts...>;
+overloaded(Ts...) -> overloaded<Ts...>;
 
 #ifdef __ANDROID__
 nn::GeneralResult<common::NativeHandle> aidlHandleFromNativeHandle(
@@ -1190,4 +1190,8 @@
     return std::vector<int32_t>(vec.begin(), vec.end());
 }
 
+std::vector<uint8_t> toVec(const std::array<uint8_t, IDevice::BYTE_SIZE_OF_CACHE_TOKEN>& token) {
+    return std::vector<uint8_t>(token.begin(), token.end());
+}
+
 }  // namespace aidl::android::hardware::neuralnetworks::utils
diff --git a/neuralnetworks/aidl/utils/src/Device.cpp b/neuralnetworks/aidl/utils/src/Device.cpp
index f3f4fdb..b64a40d 100644
--- a/neuralnetworks/aidl/utils/src/Device.cpp
+++ b/neuralnetworks/aidl/utils/src/Device.cpp
@@ -229,7 +229,6 @@
     const auto aidlDeadline = NN_TRY(convert(deadline));
     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());
@@ -240,12 +239,13 @@
         const auto ret = kDevice->prepareModelWithConfig(
                 aidlModel,
                 {aidlPreference, aidlPriority, aidlDeadline, std::move(aidlModelCache),
-                 std::move(aidlDataCache), aidlToken, std::move(aidlHints),
+                 std::move(aidlDataCache), token, std::move(aidlHints),
                  std::move(aidlExtensionPrefix)},
                 cb);
         HANDLE_ASTATUS(ret) << "prepareModel failed";
         return cb->get();
     }
+    const auto aidlToken = NN_TRY(convert(token));
     const auto ret = kDevice->prepareModel(aidlModel, aidlPreference, aidlPriority, aidlDeadline,
                                            aidlModelCache, aidlDataCache, aidlToken, cb);
     HANDLE_ASTATUS(ret) << "prepareModel failed";
diff --git a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
index 33270ff..44f8ea9 100644
--- a/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
+++ b/neuralnetworks/aidl/utils/src/InvalidDevice.cpp
@@ -189,7 +189,8 @@
         }
     }
     return prepareModel(model, config.preference, config.priority, config.deadlineNs,
-                        config.modelCache, config.dataCache, config.cacheToken, callback);
+                        config.modelCache, config.dataCache, utils::toVec(config.cacheToken),
+                        callback);
 }
 
 ndk::ScopedAStatus InvalidDevice::prepareModelFromCache(