};

TEST_P(CppGradients, TestIdentityNGrad) {
  // This test is interesting because the current implementation of GradientTape
  // would return [0, 1] whereas we use build_default_zeros_grads=false here
  // so we get back [nullptr, 1].

  AbstractTensorHandlePtr x1;
  {
    AbstractTensorHandle* x1_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(

                                    /*remote_func_outputs=*/true);
}
TEST(CAPI, RemoteExecuteSilentCopiesLocalAsyncFuncOrdering) {
  // A remote input may be not ready when we start running a function. Test that
  // the function execution should wait until the remote input is ready.
  TestRemoteExecuteSilentCopiesFunc(/*async=*/true, /*remote=*/false,

      status->status = session->session->Extend(std::move(graph_def));
      if (!status->status.ok()) {
        // Contract is we always delete input_values[i].
        return false;
      }
      // Note: session->session is not modified if Extend() fails, so
      // we only set last_num_graph_nodes if it succeeds.
      session->last_num_graph_nodes = num_nodes;
    } else {
      session->graph->mu.unlock();

  // There are two ways to represent compact row-major data
  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.
  dlm_tensor->dl_tensor.strides = stride_arr->data();

  dlm_tensor->dl_tensor.byte_offset =

  // TODO(srbs): It seems like this is used only for performance optimization
  // and not for correctness. The only downside of keeping this 1 seems to be
  // that the gradient accumulation is unbounded and we will never
  // aggressively aggregate accumulated gradients to recover memory.
  // Revisit and fix.
  return 1;
}

// Consumes references to the tensors in the gradient_tensors list and returns

                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?
      if (grad_input) {
        grad_input->Ref();
      }
      grad_inputs[i] = grad_input;
    }

  gpu_options->set_allow_growth(gpu_memory_allow_growth);

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the

      TF_SetStatusFromIOError(status, errno, dst);
      return;
    }
  } else if (S_ISDIR(st.st_mode)) {
    TF_SetStatus(status, TF_FAILED_PRECONDITION, "target path is a directory");
    return;
  }

  // We cannot rename directories yet, so prevent this.
  if (stat(src, &st) != 0) {
    TF_SetStatusFromIOError(status, errno, src);
    return;
  } else if (S_ISDIR(st.st_mode)) {

TEST(CAPI, TestRemoteExecuteUpdateServerDefResourceAccessAsync) {
  TestRemoteExecuteUpdateServerDefResourceAccess(true);
}

void TestRemoteExecuteUpdateServerDefWithFailures(bool async) {
  // Fail fast on GetStatus requests so we can get errors instead of timeout
  // when updating cluster with non-exsitent worker
  tensorflow::setenv("GRPC_FAIL_FAST", "TRUE", /*overwrite=*/1);

  tensorflow::ServerDef server_def = GetServerDef(2);

// input
//
// Description:
//   tensors.
//
//   This op can be used to override the gradient for complicated functions. For
//   example, suppose y = f(x) and we wish to apply a custom function g for
//   backprop such that dx = g(dy). In Python,
//
//   ```python
//   with tf.get_default_graph().gradient_override_map(
//       {'IdentityN': 'OverrideGradientWithG'}):