// Sessions are added to this map in TF_NewSession, and removed in
  // TF_DeleteSession.
  // TF_Graph may only / must be deleted when
  //   sessions.size() == 0 && delete_requested == true
  //
  // TODO(b/74949947): mutations currently trigger a warning instead of a bad
  // status, this should be reverted when possible.
  tensorflow::gtl::FlatMap<TF_Session*, tensorflow::string> sessions
      TF_GUARDED_BY(mu);

  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,
                                   bool heavy_load_on_streaming_rpc,

// APIs for generically dealing with op attributes (e.g. when forwarding them
// through custom device implementations).
//
// TODO(allenl): Currently these are black boxes, but we should have some way to
// inspect values. This would let people e.g. copy over most attributes and then
// modify some based on their values.

// A reference to an op's name -> attribute mapping

void TF_DeleteAbstractTensor(TF_AbstractTensor*);

// TF_OutputList holds the list of TF_AbstractTensor that results from executing
// an operation, or provided to create a function.
// When executing an operation in an eager context, the expected number of
// outputs must be set beforehand with `TF_OutputListSetNumOutputs`.
typedef struct TF_OutputList TF_OutputList;
TF_OutputList* TF_NewOutputList();

// nodes. This may cause incorrectness in graph analyses, but is useful for
// simulation or scheduling.
// If aggressive_shape_inference is true, nodes are executed on the host to
// identify output values when possible and does other aggressive strategies.
// This may cause incorrectness in graph analyses, but is useful for simulation
// or scheduling.
// If include_input_tensor_values is true, the values of constant

class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

  // (1) If type == TF_ATTR_STRING and is_list == 0
  //     then total_size is the byte size of the string
  //     valued attribute.
  // (2) If type == TF_ATTR_STRING and is_list == 1
  //     then total_size is the cumulative byte size
  //     of all the strings in the list.
  // (3) If type == TF_ATTR_SHAPE and is_list == 0
  //     then total_size is the number of dimensions
  //     of the shape valued attribute, or -1

// Returns a new thread that is running work_func and is identified
// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//
// Caller takes ownership of the result and must call TF_JoinThread on it
// eventually.

// auto* streamz = tensorflow::monitoring::Counter<1>::New("name",
// "description", "label");
// and then incremented that counter for various values of label:
// streamz->GetCell("label-value")->IncrementBy(1);
//
// The test code can then read and test the value of that counter:
//
// auto* reader = TFE_MonitoringNewCounterReader("name");
// test();