EXPECT_EQ(attr_value.i(), 2);

  // Placeholder oper now has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{feed, 0}));
  TF_Input feed_port;
  EXPECT_EQ(1, TF_OperationOutputConsumers(TF_Output{feed, 0}, &feed_port, 1));
  EXPECT_EQ(add, feed_port.oper);
  EXPECT_EQ(0, feed_port.index);

  // The scalar const oper also has a consumer.
  ASSERT_EQ(1, TF_OperationOutputNumConsumers(TF_Output{three, 0}));

                              desc->colocation_constraints.end()));
    }
    status->status = desc->node_builder.Finalize(&desc->graph->graph, &ret,
                                                 /*consume=*/true);

    if (status->status.ok()) {
      // Run shape inference function for newly added node.
      status->status = desc->graph->refiner.AddNode(ret);
    }
    if (status->status.ok()) {

                         TF_Status* s);

// Creates a new TF_AbstractFunction from the current tracing states in the
// context. The provided `ctx` is consumed by this API call and deleted.
// The returned TF_AbstractFunction must be deleted by the client,
// TODO(aminim): clarify the contract on the state of the context after this
// call.

  ~TapeVSpace() override {}

  // Returns the number of elements in the gradient tensor.
  int64_t NumElements(AbstractTensorHandle* tensor) const override;

  // Consumes references to the tensors in the gradient_tensors list and returns
  // a tensor with the result.
  AbstractTensorHandle* AggregateGradients(
      gtl::ArraySlice<AbstractTensorHandle*> gradient_tensors) const override;

      absl::Span<const AbstractTensorHandle* const> tensors) const;
  // Unwatches this tensor on the tape. Mainly used for cleanup when deleting
  // eager tensors.
  void DeleteTrace(const AbstractTensorHandle*);

  // Consumes the internal state of the tape (so cannot be called more than
  // once unless the tape is persistent) and produces the gradient of the target
  // tensors with respect to the source tensors. The output gradients are used

      memcpy(buffer,
             gcs_file->buffer.data() + (offset - gcs_file->buffer_start),
             copy_size);
    }
    bool consumed_buffer_to_eof =
        offset + copy_size >= buffer_end && gcs_file->buffer_end_is_past_eof;
    if (copy_size < n && !consumed_buffer_to_eof) {
      gcs_file->buffer_start = offset + copy_size;
      gcs_file->buffer.resize(gcs_file->buffer_size);

*   Adds new `tf.data.experimental.service.register_dataset` and
    `tf.data.experimental.service.from_dataset_id` APIs to enable one process to
    register a dataset with the tf.data service, and another process to consume
    data from the dataset.
*   Adds support for dispatcher fault tolerance. To enable fault tolerance,
    configure a `work_dir` when running your dispatcher server and set

class VSpace {
 public:
  virtual ~VSpace() {}

  // Returns the number of elements in the gradient tensor.
  virtual int64_t NumElements(Gradient* tensor) const = 0;

  // Consumes references to the tensors in the gradient_tensors list and returns
  // a tensor with the result.
  virtual Gradient* AggregateGradients(
      gtl::ArraySlice<Gradient*> gradient_tensors) const = 0;

                                                      TF_Status* status);

// In this code:
//   TF_Output producer = TF_OperationInput(consumer);
// There is an edge from producer.oper's output (given by
// producer.index) to consumer.oper's input (given by consumer.index).
TF_CAPI_EXPORT extern TF_Output TF_OperationInput(TF_Input oper_in);

// Get list of all inputs of a specific operation.  `inputs` must point to