// otherwise be the name of this custom device. Ops are placed onto a custom
  // device if any of their inputs are on that custom device, but custom devices
  // are free to set a bad status in order to require explicit placement.
  void (*execute)(const TFE_Op* op, int* num_outputs,
                  TFE_TensorHandle** outputs, TF_Status* s, void* device_info);

  // Method to delete a device.

  //
  // WARNING: The destructor of this class is marked as protected to disallow
  // clients from directly destroying this object since it may manage its own
  // lifetime through ref counting. Thus clients MUST call Release() in order to
  // destroy an instance of this class.
  virtual void Release() = 0;

  // Creates an operation builder and ties it to this context.
  // The returned object can be used for setting operation's attributes,

  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule
  // blocking collective operations at a time), and avoids some dynamic
  // allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather

// order, and type) the op declaration.  For example, the "Concat" op
// has registration:
//   REGISTER_OP("Concat")
//       .Input("concat_dim: int32")
//       .Input("values: N * T")
//       .Output("output: T")
//       .Attr("N: int >= 2")
//       .Attr("T: type");
// that defines two inputs, "concat_dim" and "values" (in that order).

    /// Wait on cond_var if state is FETCHING.
    absl::CondVar cond_var;
  };

  /// \brief The block map type for the file block cache.
  ///
  /// The block map is an ordered map from Key to Block.
  typedef std::map<Key, std::shared_ptr<Block>> BlockMap;

  /// Prune the cache by removing files with expired blocks.
  void Prune() ABSL_LOCKS_EXCLUDED(mu_);

// ignored (and hence can be None). The backprop algorithm can then leverage
// this by not constructing zeros to pass for those indices.
std::unordered_map<string, std::unordered_set<int>>*
FunctionsAcceptingNoneForIndicesMap() {
  static auto* const m =
      new std::unordered_map<string, std::unordered_set<int>>({
          {"SoftmaxCrossEntropyWithLogits", {1}},
          {"SparseSoftmaxCrossEntropyWithLogits", {1}},

// for testing.

#ifndef TENSORFLOW_C_EXPERIMENTAL_GRAPPLER_GRAPPLER_INTERNAL_H_
#define TENSORFLOW_C_EXPERIMENTAL_GRAPPLER_GRAPPLER_INTERNAL_H_

#include <functional>
#include <memory>
#include <unordered_map>
#include <unordered_set>
#include <vector>

#include "tensorflow/c/c_api.h"
#include "tensorflow/c/experimental/grappler/grappler.h"
#include "tensorflow/core/framework/graph.pb.h"

==============================================================================*/

#ifndef TENSORFLOW_C_C_API_INTERNAL_H_
#define TENSORFLOW_C_C_API_INTERNAL_H_

#include <list>
#include <set>
#include <string>
#include <unordered_map>
#include <vector>

#include "tensorflow/c/c_api.h"

// clang-format off
// Required for IS_MOBILE_PLATFORM
#include "tensorflow/core/platform/platform.h"
// clang-format on