diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.cc b/tensorflow/lite/delegates/gpu/cl/arguments.cc
index b7e6b08616e..526a09f18f9 100644
--- a/tensorflow/lite/delegates/gpu/cl/arguments.cc
+++ b/tensorflow/lite/delegates/gpu/cl/arguments.cc
@@ -843,11 +843,16 @@ absl::Status Arguments::AllocateObjects(CLContext* context) {
   for (auto& t : objects_) {
     RETURN_IF_ERROR(
         t.second.descriptor->CreateGPUObject(context, &t.second.obj_ptr));
-    t.second.descriptor->Release();
   }
   return absl::OkStatus();
 }
 
+void Arguments::ReleaseCPURepresentation() {
+  for (auto& t : objects_) {
+    t.second.descriptor->Release();
+  }
+}
+
 absl::Status Arguments::AddObjectArgs() {
   for (auto& t : objects_) {
     AddGPUResources(t.first, t.second.descriptor->GetGPUResources());
diff --git a/tensorflow/lite/delegates/gpu/cl/arguments.h b/tensorflow/lite/delegates/gpu/cl/arguments.h
index 4636a06db6f..7d76b24a47a 100644
--- a/tensorflow/lite/delegates/gpu/cl/arguments.h
+++ b/tensorflow/lite/delegates/gpu/cl/arguments.h
@@ -55,6 +55,7 @@ class Arguments {
   absl::Status Merge(Arguments&& args, const std::string& postfix);
 
   absl::Status AllocateObjects(CLContext* context);
+  void ReleaseCPURepresentation();
   absl::Status TransformToCLCode(
       const DeviceInfo& device_info,
       const std::map<std::string, std::string>& linkables, std::string* code);
diff --git a/tensorflow/lite/delegates/gpu/cl/inference_context.cc b/tensorflow/lite/delegates/gpu/cl/inference_context.cc
index 9bea02c8245..b834bbfffef 100644
--- a/tensorflow/lite/delegates/gpu/cl/inference_context.cc
+++ b/tensorflow/lite/delegates/gpu/cl/inference_context.cc
@@ -182,6 +182,10 @@ absl::Status InferenceContext::InitFromGraph(
   RETURN_IF_ERROR(Compile(creation_context));
   RETURN_IF_ERROR(UpdateParams());
 
+  for (auto& node : nodes_) {
+    node.operation->args_.ReleaseCPURepresentation();
+  }
+
   TuningParameters tuning_parameters;
   tuning_parameters.queue = env->profiling_queue();
   tuning_parameters.info = &env->device().info_;
diff --git a/tensorflow/lite/delegates/gpu/cl/kernels/cl_test.cc b/tensorflow/lite/delegates/gpu/cl/kernels/cl_test.cc
index 0112241117e..efe97f9931b 100644
--- a/tensorflow/lite/delegates/gpu/cl/kernels/cl_test.cc
+++ b/tensorflow/lite/delegates/gpu/cl/kernels/cl_test.cc
@@ -55,6 +55,7 @@ absl::Status ExecuteGPUOperation(const std::vector<TensorFloat32>& src_cpu,
 
   RETURN_IF_ERROR(operation->Compile(creation_context));
   RETURN_IF_ERROR(operation->UpdateParams());
+  operation->args_.ReleaseCPURepresentation();
   RETURN_IF_ERROR(operation->AddToQueue(creation_context.queue));
   RETURN_IF_ERROR(creation_context.queue->WaitForCompletion());