*   Add delegate support for `QUANTIZED_16BIT_LSTM`.
    *   Converts hardswish subgraphs into atomic ops.
*   Add support for defaulting the value of `cycle_length` argument of
    `tf.data.Dataset.interleave` to the number of schedulable CPU cores.
*   `parallel_for`: Add converter for `MatrixDiag`.
*   Add `narrow_range` attribute to `QuantizeAndDequantizeV2` and V3.
*   Added new op: `tf.strings.unsorted_segment_join`.

  TF_DerivedResultSizeAttr num_partitions = TF_DerivedResultSizeAttr<0>;
}

def TF_DynamicStitchOp : TF_Op<"DynamicStitch", [Pure, SameVariadicOperandSize]> {
  let summary = [{
Interleave the values from the `data` tensors into a single tensor.
  }];

  let description = [{
Builds a merged tensor such that

```python
    merged[indices[m][i, ..., j], ...] = data[m][i, ..., j, ...]