The count of forward passes through the model during generation; lower NFE means faster inference but potentially lower quality.