B.reset(B_raw);
  }

  status_ = registry_.Register("MatMul", MatMulRegisterer);
  ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  bool transpose_a_vals[] = {false, false, true, true};
  bool transpose_b_vals[] = {false, true, false, true};
  float dA_vals[4][9] = {{24, 15, 6, 24, 15, 6, 24, 15, 6},
                         {18, 15, 12, 18, 15, 12, 18, 15, 12},
                         {24, 24, 24, 15, 15, 15, 6, 6, 6},

      TFE_NewContextOptions(), TFE_DeleteContextOptions);
  std::unique_ptr<TF_Buffer, decltype(&TF_DeleteBuffer)> config(
      TF_CreateConfig(
          /*xla*/ false,
          /* gpu_memory_allow_growth */ true, /* num_cpu_devices */
          2),
      TF_DeleteBuffer);
  TFE_ContextOptionsSetConfig(opts.get(), config->data, config->length,
                              status.get());

  TFE_DeleteContext(ctx);

  // TODO(b/136478427): Figure out how to correctly shut the server down.
  worker_server.release();
}

TEST(CAPI, RemoteExecuteChangeServerDef) {
  TestRemoteExecuteChangeServerDef(false);
}
TEST(CAPI, RemoteExecuteChangeServerDefAsync) {
  TestRemoteExecuteChangeServerDef(true);
}

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

    // These XLA flags are needed to trigger XLA properly from C (more generally
    // non-Python) clients. If this API is called again with `enable` set to
    // false, it is safe to keep these flag values as is.
    tensorflow::MarkForCompilationPassFlags* flags =
        tensorflow::GetMarkForCompilationPassFlags();
    flags->tf_xla_cpu_global_jit = true;

  num_retvals = 1;
  executed = false;
  TFE_Execute(matmul.get(), &retval, &num_retvals, status.get());
  EXPECT_FALSE(executed);
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);

  // Custom devices can refuse to do type-based dispatch (as hcustom1 is
  // configured to do)
  matmul.reset(MatMulOp(context.get(), hcustom1.get(), hcpu.get()));
  num_retvals = 1;
  executed = false;

  if (node_def.op() != "Placeholder" || node_def.name() != "feed") {
    return false;
  }
  bool found_dtype = false;
  bool found_shape = false;
  for (const auto& attr : node_def.attr()) {
    if (attr.first == "dtype") {
      if (attr.second.type() == tensorflow::DT_INT32) {
        found_dtype = true;
      } else {
        return false;
      }
    } else if (attr.first == "shape") {
      found_shape = true;

    if (shape_arr[i] == 0) return true;
    // Note that dimensions with size=1 can have any stride.
    if (shape_arr[i] != 1 && stride_arr[i] != expected_stride) {
      valid = false;
    }
    expected_stride *= shape_arr[i];
  }
  return valid;
}
}  // namespace

void TFE_CallDLManagedTensorDeleter(void* dlm_ptr) {

  ConfigList configs;
  // disable_model_pruning is turned off by default.
  if (tp_configs.disable_model_pruning == TF_TriState_On)
    configs.disable_model_pruning = true;
  else
    configs.disable_model_pruning = false;
  // The other configs are turned on by default.
  CONFIG_TOGGLE(implementation_selector);
  CONFIG_TOGGLE(function_optimization);
  CONFIG_TOGGLE(common_subgraph_elimination);
  CONFIG_TOGGLE(arithmetic_optimization);

    }
  }

  TF_RETURN_IF_ERROR(GradientTape::ComputeGradient(
      vspace, target_tensor_ids, source_tensor_ids, sources_that_are_targets,
      output_gradients, result, /*build_default_zeros_grads*/ false));
  return absl::OkStatus();
}

// Helper functions which delegate to `AbstractOperation`, update
// the state of the ForwardOperation and call the tape as appropriate.

  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));
  TF_InferStatically(graph_properties, true, false, false, false, status);