tensorflow::gtl::FlatSet<int64_t> sources_set(source_tensor_ids.begin(),
                                                source_tensor_ids.end());
  std::unordered_map<int64_t, TapeTensor> sources_that_are_targets;
  for (int i = 0; i < target_tensor_ids.size(); ++i) {
    int64_t target_id = target_tensor_ids[i];
    if (sources_set.find(target_id) != sources_set.end()) {
      auto tensor = targets[i];

// is for performance optimization by reusing an exiting unused op rather than
// creating a new op every time. If `raw_device_name` is `NULL` or empty, it
// does not set the device name. If it's not `NULL`, then it attempts to parse
// and set the device name. It's effectively `TFE_OpSetDevice`, but it is faster
// than separately calling it because if the existing op has the same

  auto* task_iter = static_cast<const tensorflow::CoordinatedTask*>(tasks.data);
  for (size_t i = 0; i < tasks.length; ++i) {
    task_vec[i].set_job_name(task_iter->job_name());
    task_vec[i].set_task_id(task_iter->task_id());
    ++task_iter;
  }
  auto results = coord_agent->GetTaskState(task_vec);
  if (!results.ok()) {
    status->status = results.status();
    return;
  }

      attr_name, static_cast<const char*>(value), length);
  if (!s.ok()) {
    LOG(WARNING) << "Unable to set attribute: " << attr_name;
  }
}

void TFE_OpSetAttrInt(TFE_Op* op, const char* attr_name, int64_t value) {
  auto s = tensorflow::unwrap(op)->SetAttrInt(attr_name, value);
  if (!s.ok()) {
    LOG(WARNING) << "Unable to set attribute: " << attr_name;
  }
}

tensorflow::ServerDef GetServerDef(const string& job_name, int num_tasks) {
  tensorflow::ServerDef server_def;
  server_def.set_protocol("grpc");
  server_def.set_job_name(job_name);
  server_def.set_task_index(0);
  tensorflow::ClusterDef* cluster_def = server_def.mutable_cluster();
  tensorflow::JobDef* job_def = cluster_def->add_job();
  job_def->set_name(job_name);
  for (int i = 0; i < num_tasks; i++) {

      config.mutable_graph_options()->mutable_optimizer_options();
  if (enable) {
    optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::ON_1);

    // 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 =

// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of
// targets (not all since targets which have outputs in the tape will not have
// gradients available initially).
//

  }
  Status SetAttrFunctionName(const char* attr_name, const char* value,
                             size_t length) override {
    tensorflow::NameAttrList func_name;
    func_name.set_name(string(value, value + length));
    op_->node_builder.Attr(attr_name, func_name);
    return absl::OkStatus();
  }
  Status SetAttrTensor(const char* attr_name,

        "//tensorflow/core:framework",
        "//tensorflow/core/lib/llvm_rtti",
        "//tensorflow/core/platform:errors",
        "//tensorflow/core/platform:status",
        "@com_google_absl//absl/container:flat_hash_set",
    ],
)

tf_cuda_cc_test(
    name = "gradients_test",
    size = "small",
    srcs = [
        "gradients_test.cc",
    ],
    args = ["--heap_check="],

}

void TestFunctionWithPackedInput(const bool remote) {
  tensorflow::ServerDef server_def = GetServerDef(3);

  // This server def has the task index set to 0.
  string serialized = server_def.SerializeAsString();

  server_def.set_task_index(1);
  std::unique_ptr<tensorflow::GrpcServer> worker_server1;
  ASSERT_TRUE(tensorflow::GrpcServer::Create(