}
  auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(xla::F32, {4, 16}), "x");
  ReplicaGroup group;
  group.add_replica_ids(0);
  group.add_replica_ids(1);
  XlaOp reduce_scatter =
      ReduceScatter(x, to_apply, /*scatter_dimension=*/1, /*shard_count=*/2,
                    /*replica_groups=*/{group});

  std::vector<XlaOp> tuple_values;
  tuple_values.push_back(reduce_scatter);

        op, op.getType(), op.getInput(), op.getUpdate(),
        unpacked_indices.getOutput());
    return success();
  }
};

// Converts a TF XlaReduceScatter op to ReduceScatter HLO.
class ConvertXlaReduceScatterOp
    : public OpRewritePattern<TF::XlaReduceScatterOp> {
  using OpRewritePattern::OpRewritePattern;

  LogicalResult matchAndRewrite(TF::XlaReduceScatterOp op,

}

def TF_XlaReduceScatterOp : TF_Op<"XlaReduceScatter", [Pure]> {
  let summary = "Wraps the XLA ReduceScatter operator";

  let description = [{
documented at https://www.tensorflow.org/xla/operation_semantics#reducescatter.
  }];

  let arguments = (ins

*   `tf.test`:

    *   Added `tf.test.experimental.sync_devices`, which is useful for accurately measuring performance in benchmarks.

*   `tf.experimental.dtensor`:

    *   Added experimental support to ReduceScatter fuse on GPU (NCCL).

### Bug Fixes and Other Changes

*   `tf.SavedModel`: