int16 support

[EelcoHoogendoorn]

Looking at the code, it appears as if int16 isnt currently supported.

From my currently very supoerficial understanding, it seems like a straightforward extension of all the logic performed on int8...

But then again the fact that its not there implies the authors of this package know something about the complexity of it that I dont?

Is this on the roadmap? And are there any foreseeable problems with it?

Thanks!

[ermilovmaxim]

It's not on the roadmap, since use case is unclear.

What are you planning to use it for?
What would be the benefit over using bf16 and not quantize?

[EelcoHoogendoorn]

It is my understanding that int16 is where arm helium mcu's have a nice balance of memory and compute throughput; not sure about their fp16 performance... its certainly not as heavily advertised; but if those have the same throughput then I suppose that would be good news too.