*   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`.

    <glob pattern="*.wmx"/>
  </mime-type>
  <mime-type type="video/x-ms-wvx">
    <glob pattern="*.wvx"/>
  </mime-type>

  <mime-type type="video/x-msvideo">
    <_comment>Audio Video Interleave File</_comment>
    <alias type="video/avi"/>
    <alias type="video/msvideo"/>
    <magic priority="50">
      <match value="RIFF....AVI " type="string" offset="0"
             mask="0xFFFFFFFF00000000FFFFFFFF"/>