fix

torch_xla/csrc/init_python_bindings.cpp

@@ -1335,18 +1335,17 @@ void InitXlaModuleBindings(py::module m) {
   m.def(
       "_xla_step_marker",
       [](const std::string& device, const std::vector<std::string>& devices,
-         bool wait, const std::optional<py::dict> compile_options_args) {
+         bool wait, const std::optional<py::dict> compile_options) {
         NoGilSection nogil;
-        auto& py_options = compile_options_args;
-        if (py_options.has_value()) {
+        if (compile_options.has_value()) {
           auto wrapper = lynx::CompileOptionsWrapper::GetInstance();
           auto* completion_options_proto = &wrapper->completion_options_proto;
           auto* executable_build_options =
               completion_options_proto->mutable_executable_build_options();
           executable_build_options->set_use_spmd_partitioning(
-              py_options.value()["use_spmd_partitioning"].cast<bool>());
-          int num_replicas = py_options.value()["num_replicas"].cast<int>();
-          int num_partitions = py_options.value()["num_partitions"].cast<int>();
+              compile_options.value()["use_spmd_partitioning"].cast<bool>());
+          int num_replicas = compile_options.value()["num_replicas"].cast<int>();
+          int num_partitions = compile_options.value()["num_partitions"].cast<int>();
           executable_build_options->set_num_partitions(num_partitions);
           executable_build_options->set_num_replicas(num_replicas);
           // TODO: device assignment move to pjrt_computation_client.cc
@@ -1355,7 +1354,7 @@ void InitXlaModuleBindings(py::module m) {
         StepMarker(device, devices, wait);
       },
       py::arg("device") = "", py::arg("devices"), py::arg("wait") = true,
-      py::arg("compile_options_args") = py::none());
+      py::arg("compile_options") = py::none());
   m.def("_get_stablehlo",
         [](const std::vector<at::Tensor>& tensors, const std::string& device,
            const std::vector<std::string>& devices,