// Computes the rank of the tensor handle.
  //
  // Shapes are specified via callbacks because retrieving the shape of a tensor
  // is a blocking operation for async eager; custom devices should avoid
  // retrieving shapes of tensors they wrap until the custom device tensor's
  // shape is explicitly requested where possible.
  int (*num_dims)(void* data, TF_Status* status);

  // device-specific executors have scheduled the op.
  //
  // Accepts inferred shapes for outputs (`expected_output_shapes`), which if
  // fully defined will avoid querying the shapes of the underlying
  // TensorHandles when ParallelTensor::Shape is called. This allows async
  // computation to continue without blocking.
  //
  // The return status and value is the same as `Execute`.

  // `op_type` (e.g. "Add"), the operation's inputs (`input_tensors`,
  // `input_tensor_id`, and `input_dtypes`; the latter two are somewhat
  // 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

// (1) TF_Output or TFE_TensorHandle* as arguments to TF_AddInput,
//     TF_AddInputList
// (2) TF_ColocateWith, TF_AddControlInput etc. do not make sense.
// (3) Implementation detail: Avoid use of NodeBuilder/NodeDefBuilder since
//     the additional sanity checks there seem unnecessary;
typedef struct TFE_Op TFE_Op;

TF_CAPI_EXPORT extern TFE_Op* TFE_NewOp(TFE_Context* ctx,