//       `TP_Optimizer`, `TP_OptimizerRegistrationParams`
//     * Struct that should be filled by the proper: `TF_GrapplerItem`,
//       `TF_GraphProperties`, `TF_FunctionLibraryDefinition`
//   * We use `struct_size` for version checking. It should be set both by
//     core and the plugin.
//     * For example, `TF_InitGraph` function receives
//       `TP_OptimizerRegistrationParams*` as input with `struct_size`

// 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

              delete reinterpret_cast<DataType*>(data);
            },
            nullptr),
        TF_DeleteTensor);
    // TODO(allenl): Here and when executing regular operations, we could hold
    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.
    std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> const_op(

  // is always executed during Read.
  bool IsCacheEnabled() const { return block_size_ > 0 && max_bytes_ > 0; }

  // We can not pass a lambda with capture as a function pointer to
  // `TF_StartThread`, so we have to wrap `Prune` inside a static function.
  static void PruneThread(void* param) {
    auto ram_file_block_cache = static_cast<RamFileBlockCache*>(param);

// error information in *status.
TF_CAPI_EXPORT extern void TFE_ContextOptionsSetConfig(
    TFE_ContextOptions* options, const void* proto, size_t proto_len,
    TF_Status* status);

// Controls how to act when we try to run an operation on a given device but
// some input tensors are not on that device.
// LINT.IfChange
// Note: Keep in sync with internal copy of enum in eager/context.h.
typedef enum TFE_ContextDevicePlacementPolicy {

  // Update the Eager Executor for current thread.
  virtual void SetExecutorForThread(EagerExecutor* executor) = 0;

  // Return a list of local tensorflow::Device*.
  // TODO(tfrt-devs): We shouldn't expose legacy device in this API.
  virtual std::vector<tensorflow::Device*> ListLocalTfDevices() = 0;

  // Return a list of all tensorflow::Device*.
  virtual std::vector<tensorflow::Device*> ListAllTfDevices() = 0;

//
// Below here we do the gradient algorithm. It works as follows:
//
// First we filter the tape to just the subset of operations we want to
// differentiate. In the process of doing so we count how many times each Tensor
// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output