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();

};

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,

  // 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

    // Wait until all pending nodes have completed since they may have a
    // reference to default_cancellation_manager_. We ignore the status return
    // since we already have a bad status to propagate.
    TFE_ContextAsyncWait(context, await_status.get());
    // Reset the cancellation manager on a bad status. Otherwise we'll cancel
    // all future operations.
    default_cancellation_manager_ = absl::make_unique<CancellationManager>();
  }

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {

#define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)
#define EXPECT_TF_OK(x) EXPECT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)

static const char* content = "abcdefghijklmnopqrstuvwxyz1234567890";
// We will work with content_view instead of content.
static const absl::string_view content_view = content;

namespace gcs = google::cloud::storage;

static std::string InitializeTmpDir() {

  TFE_OpSetAttrShape(op.get(), "shape", dims, num_dims, status);
  TFE_OpSetAttrString(op.get(), "container", "", 0);
  // Use the special GUID for no buffer sharing
  //
  // TODO(allenl): Should we provide a better API for this? AFAIK this is the
  // only reasonable way to make variables with no aliasing using the eager C
  // API.
  std::string no_sharing = "cd2c89b7-88b7-44c8-ad83-06c2a9158347";

  vector<AbstractTensorHandle*> theta_inputs(inputs.size());
  for (int i{}; i < inputs.size(); ++i) {
    theta_inputs[i] = inputs[i];
  }

  AbstractTensorHandle* theta =
      theta_inputs[input_index];  // parameter we are grad checking

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.