x_signature = [
          None if c not in contracting_dims else x_shape[cidx]
          for cidx, c in enumerate(x_labels)
      ]
      y_signature = [
          None if c not in contracting_dims else y_shape[cidx]
          for cidx, c in enumerate(y_labels)
      ]
    return x_shape, y_shape, bias_shape, x_signature, y_signature

  def _create_einsum_model(
      self,

    x = Conj(scope, x);
    y = Conj(scope, y);
  }

  const auto x_shape = Shape(scope, x);
  const auto y_shape = Shape(scope, y);
  Output grad_x =
      EinsumGradWrt(scope, grad, y, x_shape, x_subs, y_subs, output_subs);
  Output grad_y =
      EinsumGradWrt(scope, grad, x, y_shape, y_subs, x_subs, output_subs);

  if (!absl::StrContains(output_subs, kEllipsis)) {

//
//   Slice(op, begin, size <must be constant>) =>
//     Slice(op, begin, actual_size(op.shape(), size, begin));
//       _XlaCompileTimeConstantInputs={2}
//
// where
//
//   actual_size(op_shape, size, begin)[i] =
//     size[i] == -1 ? (op_shape[i] - size[i])
//                   : size[i]
//
// This pass, combined with jit/partially_decluster_pass, reduces the number of

  std::vector<Output> xs;
  xs.push_back(Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)));
  xs.push_back(Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)));
  auto y = Stack(scope_, xs, Stack::Axis(1));
  TensorShape y_shape({1, 2, 2, 3});
  RunTest(xs, {x_shape, x_shape}, {y}, {y_shape});
}

TEST_F(ArrayGradTest, UnstackGrad_Axis0) {
  TensorShape x_shape({4, 2, 3});

}

void* TensorInterface::Data() const {
  return tensorflow::TensorCApi::Buffer(tensor_)->data();
}

void TensorInterface::SetShape(const int64_t* dims, int num_dims) {
  tensorflow::TensorShape s;
  for (int i = 0; i < num_dims; ++i) {
    s.AddDim(dims[i]);
  }
  tensor_.set_shape(s);
}

Status TensorInterface::BitcastFrom(const TensorInterface& from, DataType type,

  // CHECK-DAG: %[[RHS:.+]] = mhlo.subtract %[[OFFSET]], %[[MUL_MEAN]] : tensor<?xf32>
  // CHECK-DAG: %[[X_SHAPE:.+]] = shape.shape_of %[[X]] : tensor<?x?x?x?xf32> -> tensor<4xindex>
  // CHECK-DAG: %[[MULTIPLIER_BCAST:.+]] = "mhlo.dynamic_broadcast_in_dim"(%[[MULTIPLIER]], %[[X_SHAPE]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<?xf32>, tensor<4xindex>) -> tensor<?x?x?x?xf32>

#define TENSORFLOW_C_TF_SHAPE_INTERNAL_H_

#include "tensorflow/c/conversion_macros.h"
#include "tensorflow/core/framework/tensor_shape.h"

typedef struct TF_Shape TF_Shape;

namespace tensorflow {

DEFINE_CONVERSION_FUNCTIONS(tensorflow::PartialTensorShape, TF_Shape);

}

                                        device_ordinal);
  for (auto& index_to_buffer : shaped_buffer.buffers()) {
    xla::Shape subshape =
        xla::ShapeUtil::GetSubshape(on_device_shape, index_to_buffer.first);
    uint64 size =
        client->backend().transfer_manager()->GetByteSizeRequirement(subshape);
    TF_ASSIGN_OR_RETURN(se::OwningDeviceMemory buffer,
                        client->backend().memory_allocator()->Allocate(

  std::vector<int64_t> in_shape{input_type.getShape().vec()};
  std::vector<int64_t> out_shape{output_type.getShape().vec()};

  // If the reshape changes the number of dimensions so it cannot be interpreted
  // as a transpose.
  if (in_shape.size() != out_shape.size()) {
    return false;
  }

  in_shape.erase(std::remove(in_shape.begin(), in_shape.end(), 1),
                 in_shape.end());

  // The internal gradient implementation needs the shape of the input image.
  // x_shape = shape(x)[1:3]
  //         = slice(shape(x), {1}, {3 - 1})
  auto x_shape = Slice(scope, Shape(scope, op.input(0)), {1}, {2});
  grad_outputs->push_back(internal::ResizeNearestNeighborGrad(
      scope, grad_inputs[0], x_shape,
      internal::ResizeNearestNeighborGrad::AlignCorners(align_corners)