Hi Faiz,
the pre-trained models were trained for RGB images, but you can feed monochrome images into them – they should do a reasonable job at compressing them, although a model specifically trained for monochrome should perform better. I wouldn't recommend feeding 16 bit images, since the pre-trained models are designed to quantize their reconstructions to 8 bit.
If you are ok with training your own models, then it should be very easy to use the models in the models/ subdirectory for both monochrome and/or 16 bit images. You would only need to make a few adjustments in the code.
Hope this helps!
Johannes