const ParallelDevice& device_;
  std::vector<TensorHandlePtr> tensors_;
  // Parallel tensors are immutable but compute their shape lazily unless it is
  // provided on construction. The optional has a value if the lazy computation
  // has been completed or the shape was provided on construction.
  mutable absl::optional<std::vector<int64_t>> shape_;
  const TF_DataType dtype_;
};

template <typename DataType>

// * Devices are not in this API.  Instead, they are created/used internally
//   and the API just provides high level controls over the number of
//   devices of each type.

#ifdef __cplusplus
extern "C" {
#endif

// --------------------------------------------------------------------------
// TF_Version returns a string describing version information of the

  // Set up distributed execution environment on local and remote tasks.
  // When `reset_context` is true, initialize new cluster context state based
  // on cluster configurations provided in `server_def`; otherwise, update
  // existing context state with the provided `server_def`. Contexts created
  // on remote tasks will be considered stale and garbage collected after
  // `keep_alive_secs` of inactivity.

// found in the provided `op`.
TF_CAPI_EXPORT extern int TFE_OpGetInputLength(TFE_Op* op,
                                               const char* input_name,
                                               TF_Status* status);

// Returns the length (number of tensors) of the output argument `output_name`
// found in the provided `op`.
TF_CAPI_EXPORT extern int TFE_OpGetOutputLength(TFE_Op* op,

// Initialize the POSIX filesystem.
//
// In general, the `TF_InitPlugin` symbol doesn't need to be exposed in a header
// file, since the plugin registration will look for the symbol in the DSO file
// that provides the filesystem functionality. However, the POSIX filesystem
// needs to be statically registered in some tests and utilities for building
// the API files at the time of creating the pip package. Hence, we need to

  // redundant but taken as arguments to avoid repeatedly fetching these values
  // between calls to ShouldRecord and Accumulator), and its outputs
  // (`output_tensors`).
  //
  // If provided, a non-null `forward_function` will be used instead of the
  // backward function (`backward_function_getter` /
  // `backward_function_deleter`) to compute jvps for this operation. If

// returned through the provided TF_OutputList.
// Any active tape will observe the effects of this execution.
void TF_ExecuteOperation(TF_AbstractOp* op, int num_inputs,
                         TF_AbstractTensor* const* inputs, TF_OutputList* o,
                         TF_Status* s);

// Creates a new TF_AbstractFunction from the current tracing states in the
// context. The provided `ctx` is consumed by this API call and deleted.

  void (*optimize_func)(void*, const TF_Buffer*, const TF_GrapplerItem*,
                        TF_Buffer*, TF_Status*);

  // [Optional]
  // Destroy function for optimizer. If Create function is provided, destroy
  // function is must.
  void (*destroy_func)(void*);
} TP_Optimizer;

#define TP_OPTIMIZER_STRUCT_SIZE TF_OFFSET_OF_END(TP_Optimizer, destroy_func)

typedef struct TP_OptimizerRegistrationParams {

// Convenience macros for asserting and handling exceptional conditions, for
// C structs, including `TF_OpKernelContext`, `TF_Status`, etc. This is analogus
// to the macros in tensorflow/core/framework/op_requires.h. This is provided
// for plugin OpKernel developer's convenience.

#define C_OPKERNELCONTEXT_REQUIRES_OK(CTX, C_STATUS, __VA_ARGS__) \
  do {                                                            \

// metadata and synchronization.
//
// This allows us to hide concrete implementations of TensorHandle from header
// files. The interface lists the common functionality that must be provided by
// any concrete implementation. However, in cases where the true concrete class
// is needed a static_cast can be applied.
class ImmediateExecutionTensorHandle : public AbstractTensorHandle {
 public: