// =============================================================================
// Public C API entry points
//
// These are only the generic entry points for the C API. This file does not
// have any visibility into the graph/eager implementation and is only providing
// C bindings to the abstract classes defined in the
// c_api_unified_experimental_internal.h header.
//

    const Graph& graph = session->graph->graph;

    const string& mutation_warning = session->graph->sessions[session];
    if (!mutation_warning.empty()) {
      // TODO(b/74949947): turn this back into an error status
      LOG(WARNING) << mutation_warning;
      session->graph->sessions[session].clear();
    }

    const auto num_nodes = graph.num_node_ids();
    if (session->last_num_graph_nodes < num_nodes) {

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteExecutor(executor);
  TF_DeleteStatus(status);
}

// Read the value of variable `var` and save it into `out_value`.
void ReadVariable(TFE_Context* ctx, TFE_TensorHandle* var,
                  TFE_TensorHandle** out_value) {
  TF_Status* status = TF_NewStatus();
  TFE_Op* op = TFE_NewOp(ctx, "ReadVariableOp", status);

// This builder is used in the eager API to build a NodeDef.
struct TF_AttrBuilder : public tensorflow::AttrBuilder {
  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

// A ParallelDevice on its own is not registered with a TFE_Context, and so has
// no device name (e.g. for `tf.device`). `NamedParallelDevice` associates a
// name with it, which lets us pack its `ParallelTensor`s into TFE_TensorHandles
// placed on the parallel device.
class NamedParallelDevice {
 public:
  NamedParallelDevice(const std::string& name,
                      std::unique_ptr<ParallelDevice> parallel_device)

        fn_body->graph.IsValidOutputTensor(node, idx),
        "Encountered while processing input ", i, " into function '", fn_name,
        "'");
    TF_RETURN_WITH_CONTEXT_IF_ERROR(ValidateNonRefOutput(node, idx),
                                    "Encountered while processing input ", i,
                                    " into function '", fn_name, "'");

    input_tensors->emplace_back(node, idx);

  TensorHandlePtr size_three(
      VectorFloatTensorHandle(size_three_value, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Try to combine these values into a single parallel tensor.
  std::array<TFE_TensorHandle*, 2> components{size_two.get(), size_three.get()};
  TensorHandlePtr combined_value = CreatePerDeviceValues(
      context.get(), components, device_name, status.get());

                 TF_Message(first_bad_status.get()));
    return result;
  }
  // For each output of the original operation, pack the per-device
  // TensorHandles we've computed into a single parallel TensorHandle.
  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {

  // before the function execution starts.
  TFE_ContextAsyncWait(ctx, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is

    ParseGCSPath(test_dir, true, &bucket, &object, status);
    if (TF_GetCode(status) != TF_OK) {
      TF_DeleteStatus(status);
      return "";
    }
    TF_DeleteStatus(status);

    // We add a random value into `test_dir` to ensures that two consecutive
    // runs are unlikely to clash.
    std::random_device rd;
    std::mt19937 gen(rd());
    std::uniform_int_distribution<> distribution;