// a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(config, ret));
  return ret;
}

TF_Buffer* TF_CreateRunOptions(unsigned char enable_full_trace) {
  tensorflow::RunOptions options;
  if (enable_full_trace) {

                          TFE_OpGetAttrs(read_op.get()),
                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),
                          "VarHandleOp", TFE_OpGetAttrs(handle_op.get()),

  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Create a new tensor_handle.
  TFE_TensorHandle* h0_task0_new = TestMatrixTensorHandle(ctx);

  // Check that copying it to the old remote device (named localhost) fails.
  TFE_TensorHandleCopyToDevice(h0_task0_new, ctx, remote_device_name, status);
  EXPECT_NE(TF_OK, TF_GetCode(status)) << TF_Message(status);

                 TF_Status* s, const char* name, TF_Operation** op,
                 bool check) {
  TF_OperationDescription* desc = TF_NewOperation(graph, "AddN", name);
  TF_Output add_inputs[2] = {{l, 0}, {r, 0}};
  TF_AddInputList(desc, add_inputs, 2);
  *op = TF_FinishOperation(desc, s);
  if (check) {
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
    ASSERT_NE(*op, nullptr);
  }
}

  TF_ASSERT_OK(cluster->Provision());

  TrivialTestGraphInputYielder fake_input(4, 1, 10, false,
                                          cluster->GetDeviceNames());
  GrapplerItem item;
  CHECK(fake_input.NextItem(&item));

  TF_Status* status = TF_NewStatus();
  TF_GraphProperties* graph_properties =
      TF_NewGraphProperties(reinterpret_cast<TF_GrapplerItem*>(&item));

namespace {

using ::tensorflow::string;

// Add the values of three variables on three different tasks.
string AddVariablesFunction() {
  tensorflow::FunctionDef def;
  CHECK(tensorflow::protobuf::TextFormat::ParseFromString(
      "    signature {"
      "      name: 'AddVariablesFunction'"
      "      input_arg {"
      "        name: 'var'"
      "        type: DT_RESOURCE"
      "      }"

  TFE_OpAddInput(op, b, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TF_DeleteStatus(status);
  TFE_OpSetAttrType(op, "T", TFE_TensorHandleDataType(a));

  return op;
}

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b) {
  TF_Status* status = TF_NewStatus();

  TFE_Op* op = TFE_NewOp(ctx, "MatMul", status);

  // Prepare some tensors for shape.
  TF_Tensor* tensor_1X6 = Int32Tensor({1, 6});
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);
  TF_Tensor* tensor_1X1X6 = Int32Tensor({1, 1, 6});
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);

  TFE_Op* reshape_op = TFE_NewOp(tfe_context_, "Reshape", status_);
  CHECK_EQ(TF_OK, TF_GetCode(status_)) << TF_Message(status_);
  TFE_OpSetAttrType(reshape_op, "T", TF_FLOAT);

  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  int num_devices = TF_DeviceListCount(devices);
  for (int d = 0; d < num_devices; ++d) {
    const char* name = TF_DeviceListName(devices, d, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_TensorHandle* copy = TFE_TensorHandleCopyToDevice(m, ctx, name, status);

  TF_VLog(1, "Successful read of %s @ %u of size: %u", path.c_str(), offset,
          read);
  stream.read(buffer, read);
  read = stream.gcount();
  if (read < buffer_size) {
    // Check stat cache to see if we encountered an interrupted read.
    tf_gcs_filesystem::GcsFileStat stat;
    if (gcs_file->stat_cache->Lookup(path, &stat)) {
      if (offset + read < stat.base.length) {