Would it be possible to mmap the model parameters, allowing us to run even larger models? How much of an performance impact would that have?