if (forward_input) {
    TFE_DeleteTensorHandle(n);
  }

  int num_retvals = 1;
  if (async) {
    // Enqueue dummy ops so we backlog async execution & actually test async.
    // This is usually unnecessary, but we've experienced the occasional test
    // failure when testing async mode with no explicit forwarding.
    for (int i = 0; i < 100000; ++i) {
      TFE_Op* add_op_dummy = AddOp(ctx, m, m);

  ASSERT_TRUE(TF_GetCode(status) == TF_OK) << TF_Message(status);
  TF_DeleteFunction(fn);

  for (bool async : {false, true, false}) {
    TFE_Executor* old_executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_Executor* executor = TFE_NewExecutor(
        /*is_async=*/async, /*enable_streaming_enqueue=*/true,
        /*in_flight_nodes_limit=*/0);
    TFE_ContextSetExecutorForThread(ctx, executor);

  TFE_DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
} TFE_ContextDevicePlacementPolicy;
// LINT.ThenChange(//tensorflow/c/eager/immediate_execution_context.h)

// Sets the default execution mode (sync/async). Note that this can be
// overridden per thread using TFE_ContextSetExecutorForThread.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetAsync(TFE_ContextOptions*,

#include <memory>

#include "tensorflow/core/common_runtime/eager/eager_executor.h"

struct TFE_Executor {
  explicit TFE_Executor(bool async, bool enable_streaming_enqueue,
                        int in_flight_nodes_limit)
      : owned_executor(new tensorflow::EagerExecutor(
            async, enable_streaming_enqueue, in_flight_nodes_limit)) {}

  explicit TFE_Executor(tensorflow::EagerExecutor* executor)

// executed in sequence. Assigning nodes to different executors allows executing
// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are

  virtual Status InputLength(const char* input_name, int* length) = 0;
  virtual Status OutputLength(const char* output_name, int* length) = 0;

  // Set stack trace to be used for potential async error reporting.
  virtual void SetStackTrace(ManagedStackTrace stack_trace) = 0;

  virtual const tensorflow::AbstractOpAttrs* GetOpAttrs() const = 0;
  virtual void AddAttrs(const AbstractOpAttrs* op_attrs) = 0;

// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the

  TFE_Execute(op, out_value, &num_retvals, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteOp(op);
  TF_DeleteStatus(status);
}

void TestRemoteExecuteChangeServerDef(bool async) {
  tensorflow::ServerDef server_def = GetServerDef(2);

  // This server def has the task index set to 0.
  string serialized = server_def.SerializeAsString();

  server_def.set_task_index(1);

// TODO(b/154564140): Move this to its own header. This requires splitting
// c_api_experimental.h
struct TFE_ContextOptions {
  TF_SessionOptions session_options;
  // true if async execution is enabled.
  bool async = false;
  TFE_ContextDevicePlacementPolicy device_placement_policy{
      TFE_DEVICE_PLACEMENT_SILENT};
  // If true, use TFRT backend
  bool use_tfrt = false;