// The devices operations run on.
  const std::vector<std::string>& underlying_devices() const {
    return underlying_devices_;
  }

  // Takes a description of a single operation being executed on the
  // ParallelDevice, and in turn runs one operation per component device with
  // its corresponding inputs from the input ParallelTensors. Wraps the

  // Adds `value` to the existing value of the variable.
  void AssignAdd(TFE_Context* context, TFE_TensorHandle* value,
                 TF_Status* status);

 private:
  // Helper for running any single-argument assignment ops (Assign, AssignAdd,
  // AssignSub, ...).
  void GeneralAssignment(const char* op_name, TFE_Context* context,
                         TFE_TensorHandle* value, TF_Status* status);

               const std::string& send_device, const std::string& recv_device,
               tensorflow::uint64 send_device_incarnation);

// Return a 1-D INT32 tensor containing a single value 1.
TFE_TensorHandle* TestAxisTensorHandle(TFE_Context* ctx);

// Return an op taking minimum of `input` long `axis` dimension.
TFE_Op* MinOp(TFE_Context* ctx, TFE_TensorHandle* input,

//       .Attr("T: type");
// that defines two inputs, "concat_dim" and "values" (in that order).
// You must use TF_AddInput() for the first input (since it takes a
// single tensor), and TF_AddInputList() for the second input (since
// it takes a list, even if you were to pass a list with a single
// tensor), as in:
//   TF_OperationDescription* desc = TF_NewOperation(graph, "Concat", "c");
//   TF_Output concat_dim_input = {...};

// 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.
TF_CAPI_EXPORT unsigned char TF_GetXlaAutoJitEnabled();

// Sets XLA's minimum cluster size. This has global effect.

// gradients available initially).
//
// Then we repeatedly pop an entry from the stack, run its backprop, and update
// the gradients of its inputs. Once we have computed all gradients for a single
// input we can mark this input as done, and this can trigger adding an entry to
// the stack if all outputs of that entry are now done.
//
// When the stack is empty we have gradients for all tensors we're interested

  ///
  /// pattern: { term }
  /// term:
  ///   '*': matches any sequence of non-'/' characters
  ///   '?': matches a single non-'/' character
  ///   '[' [ '^' ] { match-list } ']':
  ///        matches any single character (not) on the list
  ///   c: matches character c (c != '*', '?', '\\', '[')
  ///   '\\' c: matches character c
  /// character-range:

//
// The class currently only interacts with single-slice (i.e., non-partitioned)
// variables.
class CheckpointReader {
 public:
  CheckpointReader(const string& filename, TF_Status* status);

  bool HasTensor(const string& name) const;
  const string DebugString() const;

  // Returns a map from variable names to their shapes.  Slices of a partitioned
  // tensor are combined into a single entry.