diff --git a/graph_framework/cuda_context.hpp b/graph_framework/cuda_context.hpp
index 4ed324a646a6bc98c8d1344733643e860c3d6e26..d03b9f4e269daa4aaa7547647146af672ffa61fa 100644
--- a/graph_framework/cuda_context.hpp
+++ b/graph_framework/cuda_context.hpp
@@ -756,23 +756,23 @@ namespace gpu {
                 }
             }
             for (size_t i = 0, ie = outputs.size(); i < ie; i++) {
-                if (i == 0) {
-                    if (inputs.size()) {
-                        source_buffer << ", // "
-                                      << inputs[inputs.size() - 1]->get_symbol();
+                if (!used_args.contains(outputs[i].get())) {
+                    if (i == 0) {
+                        if (inputs.size()) {
+                            source_buffer << ", // "
+                                          << inputs[inputs.size() - 1]->get_symbol();
 #ifndef USE_INPUT_CACHE
 #ifdef SHOW_USE_COUNT
-                        source_buffer << " used "
-                                      << usage.at(inputs[inputs.size() - 1].get());
+                            source_buffer << " used "
+                                        << usage.at(inputs[inputs.size() - 1].get());
 #endif
 #endif
-                        source_buffer << std::endl;
+                            source_buffer << std::endl;
+                        }
+                    } else {
+                        source_buffer << "," << std::endl;
                     }
-                } else {
-                    source_buffer << "," << std::endl;
-                }
 
-                if (!used_args.contains(outputs[i].get())) {
                     source_buffer << "    ";
                     jit::add_type<T> (source_buffer);
                     source_buffer << " *  __restrict__ "
diff --git a/graph_tests/jit_test.cpp b/graph_tests/jit_test.cpp
index 1d12fa668bb8f744e4e8df51c463f9cf9785b108..7b811b929c694b570ba2a8a9c361cf0bb01ad686 100644
--- a/graph_tests/jit_test.cpp
+++ b/graph_tests/jit_test.cpp
@@ -332,7 +332,7 @@ template<jit::float_scalar T> void run_math_tests() {
     auto sin_node = graph::sin(v1);
     compile<T> ({graph::variable_cast(v1)},
                 {sin_node}, {},
-                sin_node->evaluate().at(0), 0.0);
+                sin_node->evaluate().at(0), result);
 
     auto cos_node = graph::cos(v1);
     compile<T> ({graph::variable_cast(v1)},