absl::Span<AbstractTensorHandle* const> inputs, bool use_function,
    double abs_error = 1e-2);

void CheckTensorValue(AbstractTensorHandle* t, absl::Span<const float> manuals,
                      absl::Span<const int64_t> dims, double abs_error = 1e-2);

Model BuildGradModel(Model forward, GradientRegistry registry);

}  // namespace internal
}  // namespace gradients
}  // namespace tensorflow

extern "C" {

TF_Shape* TF_NewShape() {
  return tensorflow::wrap(new tensorflow::PartialTensorShape());
}

int TF_ShapeDims(const TF_Shape* shape) {
  return tensorflow::unwrap(shape)->dims();
}

int64_t TF_ShapeDimSize(const TF_Shape* shape, int d) {
  return tensorflow::unwrap(shape)->dim_size(d);
}

void TF_DeleteShape(TF_Shape* shape) { delete tensorflow::unwrap(shape); }

  absl::StatusOr<xla::Literal> CreateRandomF32Literal(
      absl::Span<const int64_t> dims, float min = -100, float max = 100) {
    TF_ASSIGN_OR_RETURN(auto shape,
                        xla::ShapeUtil::MakeValidatedShape(xla::F32, dims));
    return xla::LiteralUtil::CreateLiteralWithGenerator<xla::F32, float>(
        shape, [this, min, max](absl::Span<const int64_t> dims) -> float {
          return absl::Uniform(bitgen_, min, max);
        });

      } else {
        const auto num_dims = tensor_shape.dim_size();
        std::unique_ptr<int64_t[]> dims(new int64_t[num_dims]);
        for (int i = 0; i < num_dims; ++i) {
          dims[i] = tensor_shape.dim(i).size();
        }
        TFE_OpSetAttrShape(op, attr_name, dims.get(), num_dims, status);
      }
    } break;
    case tensorflow::AttrValue::kFunc: {

def LowerFillOp : Pat<(TF_FillOp $dims, $value),
                      (TF_BroadcastToOp $value, $dims)>;

//===----------------------------------------------------------------------===//
// Empty op patterns.
//===----------------------------------------------------------------------===//

def LowerEmptyOp : Pat<(TF_EmptyOp:$result $dims, TrueBoolAttr:$init),
                      (TF_BroadcastToOp

                                                TF_Status* status) {
  ::tensorflow::TensorShape shape;
  auto* cc_ctx = reinterpret_cast<::tensorflow::OpKernelConstruction*>(ctx);
  ::tensorflow::Status s = cc_ctx->GetAttr(attr_name, &shape);
  ::tensorflow::Set_TF_Status_from_Status(status, s);
  size_t rank = static_cast<size_t>(shape.dims());

  if (!status->status.ok()) return;

  if (!shaped_type) return false;

  return shaped_type.hasStaticShape();
}

bool HasStaticShapeAtDims(Value value, const ArrayRef<int> dims) {
  auto shaped_type = mlir::dyn_cast<ShapedType>(value.getType());
  if (!shaped_type || !shaped_type.hasRank()) return false;

  for (auto dim : dims) {
    if (shaped_type.isDynamicDim(dim)) return false;
  }
  return true;
}

//   shape(expand_dims(t2, 2)) ==> [2, 3, 1, 5]
//   shape(expand_dims(t2, 3)) ==> [2, 3, 5, 1]
//   ```
//
//   This operation requires that:
//
//   `-1-input.dims() <= dim <= input.dims()`
//
//   This operation is related to `squeeze()`, which removes dimensions of
//   size 1.
Status ExpandDims(AbstractContext* ctx, AbstractTensorHandle* const input,

            %cst = arith.constant dense<3.0> : tensor<32xf32>
            %cmp = "tf.Less"(%read, %cst) : (tensor<32xf32>, tensor<32xf32>) -> tensor<32xi1>
            %dims = arith.constant dense<0> : tensor<1xi32>
            %reduce = "tf.All"(%cmp, %dims) {keep_dims = false} : (tensor<32xi1>, tensor<1xi32>) -> tensor<i1>
            "tf.Yield"(%reduce) : (tensor<i1>) -> ()
        }, {

// Returns the shape of the Tensor referenced by `output` in `graph`
// into `dims`. `dims` must be an array large enough to hold `num_dims`
// entries (e.g., the return value of TF_GraphGetTensorNumDims).
//
// If the number of dimensions in the shape is unknown or the shape is
// a scalar, `dims` will remain untouched. Otherwise, each element of
// `dims` will be set corresponding to the size of the dimension. An