// 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`.

//     if you want shape inference to consider the input tensors of the
//     op for shape inference.
//   - The types need not be set in `input_shapes` as it is not used.
//   - The number of `input_tensors` should be the same as the number of items
//     in `input_shapes`.
//
// The results are returned in `output_shapes` and
// `output_resource_shapes_and_types`. The caller is responsible for freeing the

TF_CAPI_EXPORT extern void TF_DeleteGraphProperties(
    TF_GraphProperties* graph_properties);

// Infer tensor shapes through abstract interpretation.
// If assume_valid_feeds is true, it can help infer shapes in the fanout of fed
// nodes. This may cause incorrectness in graph analyses, but is useful for
// simulation or scheduling.
// If aggressive_shape_inference is true, nodes are executed on the host to

// setting a shape of [-1, 2] with an existing shape [2, -1] would set
// a final shape of [2, 2] based on shape merging semantics.
//
// Returns an error into `status` if:
//   * `output` is not in `graph`.
//   * An invalid shape is being set (e.g., the shape being set
//     is incompatible with the existing shape).
TF_CAPI_EXPORT extern void TF_GraphSetTensorShape(TF_Graph* graph,

  int version = TFE_CUSTOM_DEVICE_VERSION;

  // 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);

namespace tensorflow {

// Set the shapes and types of the output's handle.
//
// The lengths of the arrays pointed to by `shapes`, `ranks`, and `types` must
// all be equal to `num_shapes_and_types`. If `ranks[i] != -1`, (i.e., if the
// rank is known), then it must be equal to the length of `shapes[i]`; if
// `ranks[i] == 1`, then `shapes[i]` may be nullptr.
//

 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.
  const TensorSliceReader::VarToShapeMap& GetVariableToShapeMap() const;

  // Returns a map from variable names to their data types.  Slices of a

  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  // Returns the device which created the handle.
  virtual const char* DeviceName(Status* status) const = 0;

TFE_Op* MatMulOp(TFE_Context* ctx, TFE_TensorHandle* a, TFE_TensorHandle* b);

// Return an identity op.
TFE_Op* IdentityOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

#include <stack>
#include <unordered_map>
#include <unordered_set>
#include <vector>

#include "tensorflow/core/config/flag_defs.h"
#include "tensorflow/core/config/flags.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/lib/gtl/array_slice.h"
#include "tensorflow/core/lib/gtl/cleanup.h"
#include "tensorflow/core/lib/gtl/flatmap.h"