//   element. For example, if you have a single image of shape `[height, width,
//   channels]`, you can make it a batch of 1 image with `expand_dims(image,
//   0)`, which will make the shape `[1, height, width, channels]`.
//
//   Other examples:
//
//   ```
//   # 't' is a tensor of shape [2]
//   shape(expand_dims(t, 0)) ==> [1, 2]
//   shape(expand_dims(t, 1)) ==> [2, 1]
//   shape(expand_dims(t, -1)) ==> [2, 1]
//

      return ret;
    }())
  }]
>;

// A derived attribute that returns the shapes of the tensors in the actual
// value pack that corresponds to the `idx`-th ODS-declared variadic operand.
// This returns a list of shapes so it is used for variadic operands that
// can have different shapes.
class TF_DerivedOperandShapeListAttr<int idx> : DerivedAttr<
  "::mlir::TF::OperandShapeRange",

# RUN: tf-mlir-translate -graphdef-to-mlir -tf-enable-shape-inference-on-import=false %s -tf-graph-as-function -tf-control-output-arrays=assign_variable | tf-mlir-translate -mlir-tf-graph-to-hlo-text -tf-input-shapes=2:2 -tf-input-data-types=DT_FLOAT,DT_FLOAT -tf-xla-input-types=parameter,resource -tf-xla-emit-return-tuple | FileCheck %s

node {
  name: "arg0"
  op: "_Arg"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {

    key: "cond"
    value {
      func {
        name: "while_cond_5"
      }
    }
  }
  attr {
    key: "output_shapes"
    value {
      list {
        shape {
        }
        shape {
        }
        shape {
        }
      }
    }
  }
  attr {
    key: "parallel_iterations"
    value {
      i: 10
    }
  }
}
library {
  function {

// 1. does not cause an error
// 2. preserves dtype and shape.
TEST_P(SavedVariableLoadingTest, LoadSavedVariableSuccessful) {
  auto& test_params = GetParam();
  DataType dtype = std::get<0>(test_params);
  TensorShape shape(std::get<1>(test_params));

  SavedVariable saved_variable;
  saved_variable.set_dtype(dtype);
  shape.AsProto(saved_variable.mutable_shape());

  std::unique_ptr<Variable> var;

# RUN: tf-mlir-translate -graphdef-to-mlir -tf-enable-shape-inference-on-import=false %s -tf-input-arrays=x -tf-input-data-types=DT_INT32 -tf-input-shapes=10 -tf-output-arrays=func_call -o - | FileCheck %s

node {
  name: "x"
  op: "Const"
  attr {
    key: "dtype"
    value {
      type: DT_INT32
    }
  }
  attr {
    key: "value"
    value {
      tensor {
        dtype: DT_INT32
        tensor_shape {
          dim {

# CHECK-NEXT:   subgraphs: [ {
# CHECK-NEXT:     tensors: [ {
# CHECK-NEXT:       shape: [ 4 ],
# CHECK-NEXT:       type: INT32,
# CHECK-NEXT:       buffer: 1,
# CHECK-NEXT:       name: "input0",
# CHECK-NEXT:       quantization: {
# CHECK-EMPTY:
# CHECK-NEXT:       },
# CHECK-NEXT:       has_rank: true
# CHECK-NEXT:     }, {
# CHECK-NEXT:       shape: [ 4 ],
# CHECK-NEXT:       type: INT32,
# CHECK-NEXT:       buffer: 2,

    // ShapeRepresentationFn). Each bundle describes how the on-host shapes of
    // a) argument and return value, for entry computations b) variables, for
    // all computations, should be represented in XLA. Parameters/return values
    // will be shaped according to the function pair, and reshaped back to/from
    // their declared shapes for computations. Must be non-empty.
    std::vector<XlaShapeLayoutHelpers::ShapeDeterminationFns>

  // shape. If a ShapedBuffer exists already (has_shaped_buffer() == true), it
  // is replaced and the managed memory deallocated.
  Status AllocateShapedBuffer(DataType dtype, const xla::Shape& on_device_shape,
                              xla::LocalClient* client, int device_ordinal);

  // Some Tensors can have complex on-device shapes, including tuple shapes. To

class TensorSpec {
 public:
  // Constructs a scalar, DT_FLOAT TensorSpec
  TensorSpec();

  TensorSpec(PartialTensorShape shape, DataType dtype);

  explicit TensorSpec(const TensorSpecProto& proto);

  const PartialTensorShape& shape() const;
  DataType dtype() const;

 private:
  PartialTensorShape shape_;
  DataType dtype_;
};

}  // namespace tensorflow