auto attrs = oper->node.attrs();
  int count = 0;
  AttrValueMap::const_iterator it;
  for (it = attrs.begin(); it != attrs.end(); it++) {
    if (count == i) {
      strncpy(output, it->first.c_str(), it->first.length());
      status->status = absl::OkStatus();
      return;
    }
    count++;
  }
  status->status = OutOfRange("Operation only has ", count,

void TF_SetTracingImplementation(const char* name, TF_Status* s) {
  tsl::Set_TF_Status_from_Status(s, SetDefaultTracingEngine(name));
}

// Creates a new TensorFlow function, it is an execution context attached to a
// given tracing context.
TF_ExecutionContext* TF_CreateFunction(const char* fn_name, TF_Status* s) {
  return wrap(CreateTracingExecutionContext(fn_name, s));
}

  if (!status->status.ok()) {
    TF_DeleteCheckpointReader(reader);
    return nullptr;
  }
  const auto& m = reader->GetVariableToDataTypeMap();
  for (auto it = m.begin(); it != m.end(); ++it)
    reader->variable_list.push_back(it->first);
  std::sort(reader->variable_list.begin(), reader->variable_list.end());
  return reader;
}

void TF_DeleteCheckpointReader(TF_CheckpointReader* reader) { delete reader; }

} PosixFile;

static void Cleanup(TF_RandomAccessFile* file) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  close(posix_file->fd);
  // This would be safe to free using `free` directly as it is only opaque.
  // However, it is better to be consistent everywhere.
  plugin_memory_free(const_cast<char*>(posix_file->filename));
  delete posix_file;
}

static int64_t Read(const TF_RandomAccessFile* file, uint64_t offset, size_t n,

  // Although, `ops::SparseSoftmaxCrossEntropyWithLogits` returns 2 tensors, the
  // second tensor isn't needed for computing gradient so we could safely drop
  // it.
  outputs[0] = loss;
  backprop->Unref();
  return absl::OkStatus();
}

Status BiasAddModel(AbstractContext* ctx,
                    absl::Span<AbstractTensorHandle* const> inputs,

  // Import it in a fresh graph.
  TF_DeleteGraph(graph);
  graph = TF_NewGraph();
  TF_ImportGraphDefOptions* opts = TF_NewImportGraphDefOptions();
  TF_GraphImportGraphDef(graph, graph_def, opts, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  TF_Operation* scalar = TF_GraphOperationByName(graph, "scalar");

  // Import it in a fresh graph with an unused input mapping.

    } else {
      // In sync mode, forwarding can't really happen since the client code will
      // have a reference count on the input tensor while the kernel is being
      // executed and thus it cannot be re-used for the return tensor.
      EXPECT_NE(orig_ptr, TF_TensorData(t));
    }
  }

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteTensorHandle(m);

  // Finalize the function by providing the returned values.
  TF_AbstractFunction* func;
  {
    // We want to return the output of both add operations, create a new list
    // and populate it.
    TF_OutputList* func_outputs = TF_NewOutputList();
    TF_OutputListPushBack(func_outputs, add_output1, s);
    ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
    TF_OutputListPushBack(func_outputs, add_output2, s);

  char* p = ops_->translate_name(filesystem_.get(), name.c_str());
  CHECK(p != nullptr) << "TranslateName(" << name << ") returned nullptr";

  std::string ret(p);
  // Since `p` is allocated by plugin, free it using plugin's method.
  plugin_memory_free_(p);
  return ret;
}

void ModularFileSystem::FlushCaches(TransactionToken* token) {
  if (ops_->flush_caches != nullptr) ops_->flush_caches(filesystem_.get());
}

  TF_RETURN_IF_ERROR(
      ops::IdentityN(ctx, inputs, absl::MakeSpan(temp_outputs), "IdentityN"));
  // Although, `ops::IdentityN` returns 2 tensors, the first tensor isn't needed
  // for computing gradient so we could safely drop it.
  outputs[0] = temp_outputs[1];
  temp_outputs[0]->Unref();
  return absl::OkStatus();
}

class CppGradients
    : public ::testing::TestWithParam<std::tuple<const char*, bool, bool>> {
 protected: