TF_DeleteOutputList(mm_outputs);
  }

  // Finalize the function by providing the returned values.
  TF_AbstractFunction* func;
  {
    // We want to return the output of both add operations and MatMul operation,
    // 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);

// Registers a custom device for use with eager execution.
//
// Eager operations may be placed on this device, e.g.  `with
// tf.device("CUSTOM"):` from Python if `device_name` for this call is
// "/job:localhost/replica:0/task:0/device:CUSTOM:0".
//
// The custom device defines copy operations for moving TensorHandles on and
// off, and an execution operation for named operations. Often execution will

enum ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.

    TF_Operation* oper, TF_Operation** control_inputs, int max_control_inputs);

// Get the number of operations that have `*oper` as a control input.
// Note that this number can change when new operations are added to
// the graph.
TF_CAPI_EXPORT extern int TF_OperationNumControlOutputs(TF_Operation* oper);

// Get the list of operations that have `*oper` as a control input.
// `control_outputs` must point to an array of length at least

};

// Map from tensor_id to internally-defined operation-id of the operation which
// produced this tensor. A value of -1 means that the tensor was directly
// watched and not the result of any operation in the tape.
using TensorTape = std::unordered_map<int64_t, int64_t>;

// Map from operation-id to tape entry.
template <typename BackwardFunction, typename TapeTensor>
using OpTape =

          exempt-pr-labels: "override-stale"
          #Limit the No. of API calls in one run default value is 30.
          operations-per-run: 1000
          days-before-issue-stale: 7
          days-before-issue-close: 7
          stale-issue-label: "stale"
          # reason for closed the issue default value is not_planned

        operation is a scalar.
    *   Fix `NNAPI` delegate failure when Axis input for reduce operation is a
        scalar.
    *   Expose option to limit the number of partitions that will be delegated
        to `NNAPI`.
    *   If a target accelerator is specified, use its feature level to determine
        operations to delegate instead of SDK version.
*   `tf.random`:

variables as part of those interactions (ex: using a string variable to build a
filesystem path), a maliciously created checkpoint might be able to change the
targets of those operations, which could result in arbitrary
read/write/executions.

### Running a TensorFlow server

TensorFlow is a platform for distributed computing, and as such there is a

  ~DeviceThread();

  // Requests that the worker thread execute the specified operation. Blocks
  // until the previously pending operation (a StartExecute without a Join) has
  // finished, if any.
  //
  // `cancellation_manager` must live until after `Join` finishes and pending
  // `is_async` operations finish. In addition to allowing the caller to cancel
  // the operation, its `StartCancel` method will be called if op execution

  TF_Graph* graph = TF_NewGraph();

  // Make a placeholder operation.
  TF_Operation* feed = Placeholder(graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Make a constant operation with the scalar "2".
  TF_Operation* two = ScalarConst(2, graph, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  // Add operation.
  TF_Operation* add = Add(feed, two, graph, s);