// proj_clip.
    attributes.push_back(
        rewriter.getNamedAttr("proj_clip", rewriter.getF32FloatAttr(0.0)));
    // will always be time_majored.
    attributes.push_back(
        rewriter.getNamedAttr("time_major", rewriter.getBoolAttr(true)));

    Value lstm_result = rewriter.create<TFL::UnidirectionalSequenceLSTMOp>(
        op->getLoc(), result_types, inputs, attributes);

  merge_outputs: bool;

  // Parameters supported by version 2:
  // If true then first dimension is sequence, otherwise batch.
  // Version 1 implementations assumed time_major to be true, so this default
  // value should never change.
  time_major: bool = true;

  // Parameters for version 3 or above.
  asymmetric_quantize_inputs:bool;
}

table ResizeBilinearOptions {
  new_height: int (deprecated);

      %cst_7, %cst_7, %cst_7, %cst_7,
      %cst_2, %cst_2,
      %cst_1, %cell_input,
      %cst_2, %cst_2, %cst_2, %cst_2) {cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", proj_clip = 0.000000e+00 : f32, time_major = false}
    :  ( tensor<1x28x28x!quant.uniform<i8:f32, 0.003:-128>>,

  auto _time_major = _o->time_major;
  auto _fused_activation_function = _o->fused_activation_function;
  auto _asymmetric_quantize_inputs = _o->asymmetric_quantize_inputs;
  return tflite::CreateSequenceRNNOptions(
      _fbb,
      _time_major,
      _fused_activation_function,
      _asymmetric_quantize_inputs);
}

      %cst_3, %cst_2,
      %recurrent_stats, %cell_stats,
      %cst_2, %cst_2, %cst_2, %cst_2) {cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", proj_clip = 0.000000e+00 : f32, time_major = false}
    : ( tensor<1x28x28xf32>,
        tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>,
        tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>, tensor<20x20xf32>,

[[VAL_3]], [[VAL_3]], [[VAL_3]], [[VAL_3]], [[VAL_3]], [[VAL_5]], [[VAL_5]], [[VAL_4]], [[VAL_4]], [[VAL_5]], [[VAL_5]], [[VAL_5]], [[VAL_5]]) <{cell_clip = 0.000000e+00 : f32, fused_activation_function = "TANH", proj_clip = 0.000000e+00 : f32, time_major = true}> : (tensor<28x1x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x28xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16x16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>, tensor<16xf32>,...

    ConfinedAttr<DefaultValuedOptionalAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$cell_clip,
    ConfinedAttr<DefaultValuedOptionalAttr<F32Attr, "0.0f">, [TFL_FloatNonNegative]>:$proj_clip,
    BoolAttr:$time_major,
    // Used in post-training dynamic range quantization. If the value is true,
    // input activations are asymmetrically quantized.
    OptionalAttr<BoolAttr>:$asymmetric_quantize_inputs,

    return success();
  }

  // Default to non-time_major.
  std::optional<mlir::NamedAttribute> time_major_attr =
      attr.getNamed("time_major");
  bool time_majored =
      time_major_attr ? time_major_attr->getValue().cast<BoolAttr>().getValue()
                      : false;

  int64_t batch =
      time_majored ? input_type.getDimSize(1) : input_type.getDimSize(0);
  int64_t time =