auto dcast = b.create<quantfork::DequantizeCastOp>(
          dq.getLoc(), dq.getOutput().getType(), dq.getInput());
      dq.getOutput().replaceAllUsesWith(dcast);
      dq.erase();
    } else if (auto q = llvm::dyn_cast<QuantizeOp>(op)) {
      auto qcast = b.create<quantfork::QuantizeCastOp>(
          q.getLoc(), q.getOutput().getType(), q.getInput());
      q.getOutput().replaceAllUsesWith(qcast);

      auto new_read_variable_op =
          builder.create<ReadVariableOp>(read_variable_op.getLoc(), ref_qtype,
                                         read_variable_op.getResourceId());
      auto new_dq_op = builder.create<DequantizeOp>(
          read_variable_op.getLoc(), read_variable_op.getResult().getType(),
          new_read_variable_op.getResult());

        req_op.getLoc(), func_op.getResultTypes()[0], req_op.getOperand());
    if (func_name.contains("_relu6_")) {
      auto min = rewriter.create<mlir::stablehlo::ConstantOp>(
          req_op.getLoc(), rewriter.getF32FloatAttr(0));
      auto max = rewriter.create<mlir::stablehlo::ConstantOp>(
          req_op.getLoc(), rewriter.getF32FloatAttr(6));

  auto erf_op = rewriter.create<chlo::ErfOp>(mul_op.getLoc(), mul_op);
  auto add_op =
      rewriter.create<stablehlo::AddOp>(erf_op.getLoc(), erf_op, one_cst);
  auto lhs_mul_op = rewriter.create<stablehlo::MulOp>(
      half_cst.getLoc(), new_func.getArguments()[0], half_cst);
  auto output_mul_op = rewriter.create<stablehlo::MulOp>(lhs_mul_op.getLoc(),

      auto tf_reduce_op = rewriter.create<BooleanReduce>(
          reduce_op.getLoc(), reduce_op->getResult(0).getType(), operand,
          reduction_indices,
          /*keep_dim=*/rewriter.getBoolAttr(false));
      auto tf_argreduce_op = rewriter.create<ArgReduce>(
          reduce_op.getLoc(), reduce_op->getResult(1).getType(), operand,
          reduction_indices);

  builder.insert(var_handle_op_in_init);
  auto const_op = builder.create<mlir::arith::ConstantOp>(
      session_init_func.getLoc(), tensor_attr.getType(), tensor_attr);

  builder.create<TF::AssignVariableOp>(
      session_init_func.getLoc(), llvm::ArrayRef<mlir::Type>{},
      llvm::ArrayRef<mlir::Value>{var_handle_op_in_init->getResult(0),
                                  const_op.getResult()});
}

    Type output_type = bmm_op.getResult().getType();
    auto no_input = rewriter.create<TFL::NoValueOp>(
        bmm_op->getLoc(), rewriter.getNoneType(), rewriter.getUnitAttr());
    auto fc_op = rewriter.create<TFL::FullyConnectedOp>(
        bmm_op->getLoc(), ArrayRef<Type>{output_type},
        /*input=*/output_lhs, /*filter=*/output_rhs, /*bias=*/no_input,

      replacer.create<mlir::func::ReturnOp>(yield_op.getLoc(),
                                            yield_op.getOperands());
      yield_op.erase();
    }

    // Remap the captured operands in the (former) island block with newly
    // created entry block arguments in the function body.
    {
      Block &entry_block = outlined_func.getBody().front();
      auto loc = outlined_func.getLoc();
      for (Value operand : operands) {

        mlir::cast<TensorType>(input.getType()), kNchwToNhwcPermutation);

    auto input_transpose_op = rewriter.create<mlir::stablehlo::TransposeOp>(
        op.getLoc(), /*resultType0=*/new_input_tensor_type, /*operand=*/input,
        rewriter.getDenseI64ArrayAttr(kNchwToNhwcPermutation));

    // Transpose the filter tensor: [o, i, 0, 1] => [0, 1, i, o]
    Value filter = op->getOperand(1);

  OpBuilder builder(func_.getBody());
  auto op = builder.create<mlir::TFL::NonMaxSuppressionV4Op>(
      func_.getLoc(), output_type0, output_type1, boxes, scores,
      max_output_size, iou_threshold, score_threshold);

  builder.create<mlir::func::ReturnOp>(func_.getLoc(), op.getResults());
}

LogicalResult ConvertNMSPaddedFunc::VerifySignature() {
  // Verify high-level function signature.