// quantize pair around the op.
void ConvertQuantizedOpToFloat(func::FuncOp func, OpBuilder* builder);

// This will optimize the quantized ops -> float graph.
void OptimizeQuantizedOpToFloat(func::FuncOp func, MLIRContext* context);

}  // namespace tac
}  // namespace TFL
}  // namespace mlir

    if (!quant_op.getResult().use_empty()) return failure();

    rewriter.eraseOp(quant_op);
    return success();
  }
};

void OptimizeQuantizedOpToFloat(func::FuncOp func, MLIRContext* context) {
  RewritePatternSet patterns(func.getContext());
  patterns
      .add<FoldQuantizedI32ToFloat, FoldQuantizeDequantize, RemoveUnusedQuant>(
          context);

      target_device_inference_type.inference_type == FLOAT) {
    OpBuilder cloned_func_builder(cloned_func);
    ConvertQuantizedOpToFloat(cloned_func, &cloned_func_builder);
    OptimizeQuantizedOpToFloat(cloned_func, &getContext());
  }

  Optimize(cloned_func, target_device_inference_type.hardware);

  // Set device for each op.
  cloned_func.walk([&](Operation* op) {