auto seq_lengths = op.input(1);
  int batch_dim;
  TF_RETURN_IF_ERROR(GetNodeAttr(op.node()->attrs(), "batch_dim", &batch_dim));
  int seq_dim;
  TF_RETURN_IF_ERROR(GetNodeAttr(op.node()->attrs(), "seq_dim", &seq_dim));
  grad_outputs->push_back(
      ReverseSequence(scope, grad_inputs[0], seq_lengths, seq_dim,
                      ReverseSequence::BatchDim(batch_dim)));
  grad_outputs->push_back(NoGradient());
  return scope.status();
}

          ConstantAttr<RankedI64ElementsAttr<[]>,"static_cast<int64_t>(0)">))>;
def LegalizeReverseSequence : Pat<
  (TF_ReverseSequenceOp $input, $seq_lengths, $seq_dim, $batch_dim),
  (TFL_ReverseSequenceOp $input, $seq_lengths,
      (convertIntAttrTo32Bit $seq_dim), (convertIntAttrTo32Bit $batch_dim))>;
def LegalizeRound : Pat<(TF_RoundOp $arg), (TFL_RoundOp $arg)>;
def LegalizeRsqrt : Pat<(TF_RsqrtOp $arg), (TFL_RsqrtOp $arg)>;

  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));
  auto seq_lengths = Const(scope_, {1, 2, 3, 4, 5});
  // batch_dim defaults to 0.
  auto y = ReverseSequence(scope_, x, seq_lengths, /* seq_dim */ 2);
  RunTest(x, shape, y, shape);
}

TEST_F(ArrayGradTest, ReverseGrad) {
  TensorShape shape({5, 2, 5});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(shape));

the dimension `seq_dim`.

The elements of `seq_lengths` must obey `seq_lengths[i] <= input.dims[seq_dim]`,
and `seq_lengths` must be a vector of length `input.dims[batch_dim]`.

The output slice `i` along dimension `batch_dim` is then given by input
slice `i`, with the first `seq_lengths[i]` slices along dimension
`seq_dim` reversed.
  }];

  let arguments = (ins

  func.func @reverse_sequence(%arg0: tensor<4x2x3x1x1xi32>, %arg1: tensor<3xi32>) -> tensor<4x2x3x1x1xi32> {
    // CHECK-NOT: tf.ReverseSequence
    %0 = "tf.ReverseSequence"(%arg0, %arg1) {batch_dim = 2 : i64, seq_dim = 0 : i64}: (tensor<4x2x3x1x1xi32>, tensor<3xi32>) -> tensor<4x2x3x1x1xi32>
    func.return %0 : tensor<4x2x3x1x1xi32>
  }

  // CHECK-LABEL: mirror_pad

  auto _seq_dim = _o->seq_dim;
  auto _batch_dim = _o->batch_dim;
  return tflite::CreateReverseSequenceOptions(
      _fbb,
      _seq_dim,
      _batch_dim);
}

inline MatrixDiagOptionsT *MatrixDiagOptions::UnPack(const ::flatbuffers::resolver_function_t *_resolver) const {

  %0 = "tf.ReverseSequence"(%arg0, %arg1) {seq_dim = 0 : i64, batch_dim = 0 : i64}: (tensor<2x3xf32>, tensor<2xi32>) -> tensor<2x3xf32>
  func.return %0: tensor<2x3xf32>

// CHECK-LABEL: ReverseSequence
// CHECK:  "tfl.reverse_sequence"(%arg0, %arg1) <{batch_dim = 0 : i32, seq_dim = 0 : i32}> : (tensor<2x3xf32>, tensor<2xi32>) -> tensor<2x3xf32>
}

the dimension `seq_dim`.

The elements of `seq_lengths` must obey `seq_lengths[i] <= input.dims[seq_dim]`,
and `seq_lengths` must be a vector of length `input.dims[batch_dim]`.

The output slice `i` along dimension `batch_dim` is then given by input
slice `i`, with the first `seq_lengths[i]` slices along dimension
`seq_dim` reversed.

For example:

```

  if (sequencer_ == nullptr) {
    NodeDef seq_def;
    // TODO(shikharagarwal): What source node should we use for errors?
    NodeDefBuilder builder(absl::StrCat(subgraph_name, "_sequencer"), "NoOp");
    builder.Attr(kXlaHostTransferSequencerAttr, subgraph_name);
    builder.Device(device_);
    Status s = builder.Finalize(&seq_def);
    if (!s.ok()) return s;

`tf.reduce_min`: `reduction_indices` becomes `axis` * `tf.reduce_prod`:
`reduction_indices` becomes `axis` * `tf.reduce_sum`: `reduction_indices`
becomes `axis` * `tf.reverse_sequence`: `batch_dim` becomes `batch_axis`,
`seq_dim` becomes `seq_axis` * `tf.sparse_concat`: `concat_dim` becomes `axis` *
`tf.sparse_reduce_sum`: `reduction_axes` becomes `axis` *
`tf.sparse_reduce_sum_sparse`: `reduction_axes` becomes `axis` *