// The "truth" labels are at index 1.
  auto softmax_grad = op.output(1);

  // The loss is the output at index 0, and backprop is the output at index 1.
  auto grad_loss = grad_inputs[0];
  auto grad_grad = grad_inputs[1];

  auto grad = BroadcastMul(scope, grad_loss, softmax_grad);
  if (!IsZero(scope, grad_grad)) {
    std::vector<int> axis;
    auto logits_softmax = Softmax(scope, logits);

*   Fixes a security vulnerability caused by lack of validation in
    `tf.raw_ops.DataFormatVecPermute` and `tf.raw_ops.DataFormatDimMap`
    ([CVE-2020-26267](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2020-26267))
*   Fixes a vulnerability caused by attempting to write to immutable memory
    region in `tf.raw_ops.ImmutableConst`
    ([CVE-2020-26268](https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2020-26268)

  int axis;
  TF_RETURN_IF_ERROR(GetNodeAttr(op.node()->attrs(), "axis", &axis));

  grad_outputs->reserve(N);
  auto grad_op = Unstack(scope, grad_inputs[0], N, Unstack::Axis(axis));
  for (const Output& o : grad_op.output) {
    grad_outputs->emplace_back(o);
  }
  return scope.status();
}
REGISTER_GRADIENT_OP("Pack", PackGrad);

    // safe.
    if (auto tensor_array_v3_op = dyn_cast<TF::TensorArrayV3Op>(tf_op)) {
      // The "flow" in TensorArrayV3 is always a scalar float tensor.
      // https://www.tensorflow.org/api_docs/python/tf/raw_ops/TensorArrayWriteV3
      Value flow = tensor_array_v3_op.getFlow();
      Type scalar_f32_tensor_type =
          RankedTensorType::get(/*shape=*/{}, rewriter.getF32Type());
      flow.setType(scalar_f32_tensor_type);

  // possible terminator.
  return std::next(block->begin()) == block->end() ||
         std::next(block->begin(), 2) == block->end();
}

// This pattern matches GraphOps with only one FetchOp (empty) and remaps the
// results of the GraphOp to the operands of the FetchOp.
struct DropEmptyGraph : public OpRewritePattern<GraphOp> {
  using OpRewritePattern<GraphOp>::OpRewritePattern;

    # TODO(mdan): Index more efficiently. Could do a name check instead.
    if any(v is value for v in AG_MODULE.__dict__.values()):
      return {TFRTypes.AG_BUILTIN_FUNC}
    if getattr(value, '__name__', None) == 'tensorflow.raw_ops':
      return {types.ModuleType}
    if hasattr(value, '__module__'):
      if isinstance(value, dtypes.DType):
        return {TFRTypes.ATTR}

  def model_fn(args):
    # do something with dynamic tensor

  @function.Defun(capture_resource_var_by_value=False)
  def tpu_subgraph():
      return tf.tpu.rewrite(model_fn, args)

  return tf.raw_ops.TPUPartitionedCall(
      args=tpu_subgraph.captured_inputs,
      Tout=[o.type for o in tpu_subgraph.definition.signature.output_arg],
      f=tpu_subgraph,
      device_ordinal=[0])

elements in `x` must be complex numbers of the form \\(a + bj\\). The absolute
value is computed as \\( \sqrt{a^2 + b^2}\\).

For example:

>>> x = tf.complex(3.0, 4.0)
>>> print((tf.raw_ops.ComplexAbs(x=x, Tout=tf.dtypes.float32, name=None)).numpy())
5.0
  }];

  let arguments = (ins
    TensorOf<[TF_Complex128, TF_Complex64]>:$x
  );

  let results = (outs

// AllToAll op legalizations.
//===----------------------------------------------------------------------===//

// -----

// CHECK-LABEL: func @alltoall_basic
// See https://www.tensorflow.org/api_docs/python/tf/raw_ops/AllToAll
func.func @alltoall_basic(%input: tensor<1x2xf32>) -> tensor<2x1xf32> {
  %group_assignment = "tf.Const" () {
    value = dense<[[0, 1]]> : tensor<1x2xi32>
  } : () -> tensor<1x2xi32>