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.

      if (iter == input_nodes.end()) {
        // This node is not referenced in inputs. Add it to the body.
        body_nodes->push_back(node);
      } else {
        // This node is referenced in inputs. Currently, we place an
        // artificial restriction and require that when num_opers=-1, such
        // nodes must have a single output.
        if (node->num_outputs() != 1) {
          return InvalidArgument(

  TFE_ContextSetServerDef(ctx, 0, serialized.data(), serialized.size(), status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Use large matrices so that RPCs don't return before we get a chance
  // to call TFE_DeleteContext.
  TFE_TensorHandle* h0_task0 = TestMatrixTensorHandle100x100(ctx);
  TFE_TensorHandle* h1_task0 = TestMatrixTensorHandle100x100(ctx);
  const char remote_device_name[] =

}

TEST(PARALLEL_DEVICE, TestCollectiveSync) { TestCollective(/*async=*/false); }

// Note that ops on the parallel device currently don't execute
// asynchronously. The test is just that we don't get deadlocks.
TEST(PARALLEL_DEVICE, TestCollectiveAsync) { TestCollective(/*async=*/true); }

void RegisterCollectiveMulFunction(TFE_Context* context,

  return OkStatus();
}

// Validates the random access file operations supplied by the plugin.
static Status ValidateHelper(const TF_RandomAccessFileOps* ops) {
  if (ops == nullptr) {
    // We allow filesystems where files can only be written to (from TF code)
    return OkStatus();
  }

  if (ops->cleanup == nullptr)
    return errors::FailedPrecondition(

  int64_t read;
  auto content_length = stream.headers().find("content-length");
  if (content_length == stream.headers().end()) {
    // When we read a file with offset that is bigger than the actual file size.
    // GCS will return an empty header (e.g no `content-length` header). In this
    // case, we will set read to `0` and continue.
    read = 0;
  } else if (!absl::SimpleAtoi(content_length->second, &read)) {

    const char* fn_name = "test_fn";
    core::RefCountPtr<AbstractFunction> scoped_func;
    // Returning null tensors from a tf.function is not supported, so we keep
    // track of indices in the model's outputs are nullptr in this set.
    // The FunctionDef only outputs the non-null tensors. We later pad the
    // function op outputs to have nullptrs at the `null_indices`.
    absl::flat_hash_set<int> null_indices;
    {