// live until `Join` finishes. If `is_async=true` it must live until `Join` is
  // followed by `TFE_ContextAsyncWait` to clear pending operations. It will be
  // used to cancel all other operations if any fails.
  //
  // Set step_id to configure the step id used for rendezvous creation. step id
  // of value -1 is reserved for global rendezvous and should not be set here.

// Execute the operation defined by 'op' and return handles to computed
// tensors in `retvals`.
//
// 'retvals' must point to a pre-allocated array of TFE_TensorHandle* and
// '*num_retvals' should be set to the size of this array. It is an error if
// the size of 'retvals' is less than the number of outputs. This call sets
// *num_retvals to the number of outputs.
//

// Tensor or IndexedSlices or None, which here we map to nullptr. Gradients need
// to allow their size to be computed and they need to be passable to a backward
// function and deleted (as the backprop code creates lots of gradients the user
// is not interested in).
//
// BackwardFunction needs to be a closure which stores intermediate activations

//
// The API here is subject to changes in the future.
// --------------------------------------------------------------------------

#ifdef __cplusplus
extern "C" {
#endif

// When `enable` is true, set
// tensorflow.ConfigProto.OptimizerOptions.global_jit_level to ON_1, and also
// set XLA flag values to prepare for XLA compilation. Otherwise set
// global_jit_level to OFF.
//

  // signifying to pin to the current custom device. Returns false
  // to pin to the physical device.
  //
  // This function is guaranteed to be called only when all of the custom-device
  // inputs are on this device.
  bool (*shall_pin_to_this_device)(const TFE_Op* op, TF_Status* s) = nullptr;
} TFE_CustomDevice;

// Registers a custom device for use with eager execution.
//
// Eager operations may be placed on this device, e.g.  `with

} TF_TriState;

// TF_GrapplerItem represents a combination of a graph, one of more fetch nodes,
// and potentially a set of nodes to feed.
typedef struct TF_GrapplerItem TF_GrapplerItem;

// Flags indicating whether existing optimizers should be turned off.
// It's optional for plugin to set functions to return true/false. If not
// set, proper uses configuration set by user.
typedef struct TP_OptimizerConfigs {
  size_t struct_size;

  ///
  /// In order to prevent deadlocks, never grab the block-cache-wide mu_ lock
  /// AFTER grabbing any block's mu lock. It is safe to grab mu without locking
  /// mu_.
  struct Block {
    /// The block data.
    std::vector<char> data;
    /// A list iterator pointing to the block's position in the LRU list.
    std::list<Key>::iterator lru_iterator;

  // Create a handle to wrap and manage a Tensor
  virtual ImmediateExecutionTensorHandle* CreateLocalHandle(
      AbstractTensorInterface* t) = 0;
  // Copy the handle to another device.
  virtual ImmediateExecutionTensorHandle* CopyTensorHandleToDevice(
      ImmediateExecutionTensorHandle* handle, const char* device_name,
      Status* status) = 0;

  // Create an operation to perform op execution