/// max item age (in seconds) and max entries.
///
/// This class is thread safe.
template <typename T>
class ExpiringLRUCache {
 public:
  /// A `max_age` of 0 means that nothing is cached. A `max_entries` of 0 means
  /// that there is no limit on the number of entries in the cache (however, if
  /// `max_age` is also 0, the cache will not be populated).
  ExpiringLRUCache(uint64_t max_age, size_t max_entries,

// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be
// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
void AllocateParallelDevice(const char* device_name,
                            const char* const* underlying_devices,

// 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.

//
// Conventions:
// * We use the prefix TF_ for everything in the API.
// * Objects are always passed around as pointers to opaque structs
//   and these structs are allocated/deallocated via the API.
// * TF_Status holds error information.  It is an object type
//   and therefore is passed around as a pointer to an opaque
//   struct as mentioned above.
// * Every call that has a TF_Status* argument clears it on success

    unsigned char enable);
TF_CAPI_EXPORT unsigned char TF_SetTfXlaCpuGlobalJit(unsigned char enable);

// Sets XLA's auto jit mode according to the specified string, which is parsed
// as if passed in XLA_FLAGS. This has global effect.
TF_CAPI_EXPORT void TF_SetXlaAutoJitMode(const char* mode);

// Returns whether the single GPU or general XLA auto jit optimizations are
// enabled through MarkForCompilationPassFlags.

  // a tensor with the result.
  virtual Gradient* AggregateGradients(
      gtl::ArraySlice<Gradient*> gradient_tensors) const = 0;

  // Calls the passed-in backward function.
  //
  // `unneeded_gradients` contains sorted list of input indices for which a
  // gradient is not required.
  virtual Status CallBackwardFunction(

                                      int64_t file_signature)
      ABSL_LOCKS_EXCLUDED(mu_);

  /// Remove all cached blocks for `filename`.
  void RemoveFile(const std::string& filename) ABSL_LOCKS_EXCLUDED(mu_);

  /// Remove all cached data.
  void Flush() ABSL_LOCKS_EXCLUDED(mu_);

  /// Accessors for cache parameters.
  size_t block_size() const { return block_size_; }

#endif

// TF_Device is a C wrapper to the C++ TF Device class. This is to be passed
// through TF_OpKernelContext, and is opaque to plugin.
typedef struct TF_Device TF_Device;

typedef struct TF_VariableInfo TF_VariableInfo;

// Returns a `TF_Device` pointer, which actually points to a C++ `Device`.
// Currently we only allow `NextPluggableDevice` to be casted as `TF_Device`,
// but in theory every this is a C API for every kind of device.