// completed. This is only valid on local TFE_TensorHandles. The pointer
// returned will be on the device in which the TFE_TensorHandle resides (so e.g.
// for a GPU tensor this will return a pointer to GPU memory). The pointer is
// only guaranteed to be valid until TFE_DeleteTensorHandle is called on this
// TensorHandle. Only supports POD data types.

class ParallelTensor;
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);

// gradient function can accept None values. e.g. FusedBatchNorm outputs 5
// values and hence receives 5 gradient values during backprop. However the
// gradient function uses only the first of those values and ignores the rest.
// The entry, "FusedBatchNorm": [1, 2, 3, 4], indicates that only the gradient
// corresponding to index 0 is used, and the gradient values at indices 1-4 are
// ignored (and hence can be None). The backprop algorithm can then leverage

// this to be called.
TF_CAPI_EXPORT void TF_InitMain(const char* usage, int* argc, char*** argv);

// Platform-specific implementation to return an unused port. (This should used
// in tests only.)
TF_CAPI_EXPORT int TF_PickUnusedPortOrDie(void);

// Fast path method that makes constructing a single scalar tensor require less
// overhead and copies.

//
// Note that TFE_OpGetFlatInputCount and TFE_OpGetFlatInput operate on a flat
// sequence of inputs, unlike TFE_OpGetInputLength (for getting the length of a
// particular named input list, which may only be part of the op's inputs).
TF_CAPI_EXPORT extern int TFE_OpGetFlatInputCount(const TFE_Op* op,
                                                  TF_Status* status);
// Returns a borrowed reference to one of `op`'s inputs. Use

  ///
  /// Thread safety:
  /// The iterator and timestamp fields should only be accessed while holding
  /// the block-cache-wide mu_ instance variable. The state variable should only
  /// be accessed while holding the Block's mu lock. The data vector should only
  /// be accessed after state == FINISHED, and it should never be modified.
  ///

  virtual void StartStep() = 0;
  // Destroy the step resource container for a training step.
  virtual void EndStep() = 0;

  // Return the Eager Executor for current thread. Please note that Eager
  // Executor is only used in current TF but not in TFRT.
  virtual EagerExecutor& Executor() = 0;
  // Update the Eager Executor for current thread.
  virtual void SetExecutorForThread(EagerExecutor* executor) = 0;