control_output_nodes.reserve(ncontrol_outputs);
  for (int i = 0; i < ncontrol_outputs; ++i) {
    control_output_nodes.push_back(&control_outputs[i]->node);
  }

  // Do the actual function creation.
  DCHECK(append_hash_to_fn_name <= 1);
  tensorflow::FunctionDef fdef;
  status->status = tensorflow::GraphToFunctionDef(
      fn_body->graph, fn_name, append_hash_to_fn_name != 0,

  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph

#ifdef __cplusplus
extern "C" {
#endif

// Resets `op_to_reset` with `op_or_function_name` and `raw_device_name`. This
// is for performance optimization by reusing an exiting unused op rather than
// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse
// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster

    ]),
)

# Check that c_api_no_xla does not depend on xla.
check_deps(
    name = "c_api_no_xla_check_deps",
    disallowed_deps = ["//tensorflow/compiler/jit:xla_kernel_creator"],
    deps = [":c_api_no_xla"],
)

tf_cuda_library(
    name = "c_api_no_xla",
    srcs = [
        "c_api.cc",
        "c_api_function.cc",
    ],
    hdrs = [
        "c_api.h",

    std::vector<TensorHandlePtr> components, absl::Span<const int64_t> shape,
    TF_Status* status) {
  if (components.empty()) {
    TF_SetStatus(status, TF_INTERNAL,
                 "No components are provide for creating a ParallelTensor");
    return nullptr;
  }
  TFE_TensorHandleGetStatus(components[0].get(), status);
  if (!status->status.ok()) {
    return nullptr;
  }