Node* node = graph->graph.FindNodeId(*pos);
    // FindNodeId() returns nullptr for nodes that have been deleted.
    // We aren't currently allowing nodes to be deleted, but it is safer
    // to still check.
    if (node != nullptr) return ToOperation(node);
    *pos += 1;
  }

  // No more nodes.
  return nullptr;
}

void TF_GraphToGraphDef(TF_Graph* graph, TF_Buffer* output_graph_def,

                          /*expected_max_outputs=*/1, status.get());
  ASSERT_FALSE(TF_GetCode(status.get()) == TF_OK);
  TF_SetStatus(status.get(), TF_OK, "");

  // Check that ops still run successfully on the device.
  parallel_device.Execute(context.get(), std::vector<ParallelTensor*>(),
                          "VarHandleOp", TFE_OpGetAttrs(handle_op.get()),

  // timestamp of the entry.
  cache.Insert(key, 41);
  env->SetNowSeconds(2);
  cache.Insert(key, 42);
  // 1 second after the most recent insertion, the entry is still valid.
  env->SetNowSeconds(3);
  int value = 0;
  EXPECT_TRUE(cache.Lookup(key, &value));
  EXPECT_EQ(value, 42);
  // 2 seconds after the most recent insertion, the entry is no longer valid.
  env->SetNowSeconds(4);

  TF_EXPECT_OK(ReadCache(&cache, "", 2 * block_size, 1, &out));
  // LRU element was at offset 0.  Cache miss.
  calls.push_back(0);
  TF_EXPECT_OK(ReadCache(&cache, "", 0, 1, &out));
  // Element at 2 * block_size is still in cache, and this read should update
  // its position in the LRU list so it doesn't get evicted by the next read.
  TF_EXPECT_OK(ReadCache(&cache, "", 2 * block_size, 1, &out));

  // in the cache, and our current block is not block size, this likely means
  // we have inconsistent state within the cache. Note: it's possible some
  // incomplete reads may still go undetected.
  if (block->data.size() < block_size_) {
    Key fmax = std::make_pair(key.first, std::numeric_limits<size_t>::max());
    auto fcmp = block_map_.upper_bound(fmax);

};

int ModularFileSystemTest::rng_val_;
std::string ModularFileSystemTest::cloud_path_;
std::string ModularFileSystemTest::tmp_dir_;

// As some of the implementations might be missing, the tests should still pass
// if the returned `Status` signals the unimplemented state.
bool UnimplementedOrReturnsCode(Status actual_status, Code expected_code) {
  Code actual_code = actual_status.code();

    EXPECT_TRUE(strstr(TF_Message(status), msg) != nullptr)
        << TF_Message(status);
    // Since error is not cleared, the following copy with correct device will
    // still fail.
    TF_SetStatus(status, TF_OK, "");
    TFE_DeleteTensorHandle(retvals[0]);
    TFE_Executor* executor = TFE_ContextGetExecutorForThread(ctx);
    TFE_ExecutorWaitForAllPendingNodes(executor, status);

  EXPECT_TRUE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_DeleteTensorHandle(retval);

  // Explicit placement still forces the op onto the requested device
  matmul.reset(MatMulOp(context.get(), hcustom0.get(), hcpu.get()));
  TFE_OpSetDevice(matmul.get(), "/job:localhost/replica:0/task:0/device:CPU:0",
                  status.get());

//   * allows having newer plugins than the current core TensorFlow: the
//     additional entries in the plugin's table are just discarded;
//   * allows having older plugins than the current core TensorFlow (though
//     we are still warning users): the entries that core TensorFlow expects
//     but plugins didn't provide will be set to `nullptr` values and core
//     TensorFlow will know to not call these on behalf of users;

  }
  ~GraphOperation() override {}

 private:
  friend class GraphContext;  // For access to op_.
  TF_Graph* g_;
  std::unique_ptr<TF_OperationDescription> op_;
  // Hold `op_type` and `op_name` till both are available since we need both
  // to build a graph operation.
  string op_type_;
  const char* op_name_ = nullptr;
  // TODO(srbs): Use this.
  string device_name_;
};