Stream execution in the runtime


I am new and make my first steps in tvm. I want to deploy to opencl (and maybe metal) models via the C++ API. Since the application is working with videos I wonder if “stream execution” is supported and how to do it: Can I upload every input frame asyncronly and execute the stream? And would it be possible to hold e.g. 5 results in the GPU memory and give them as input the an another network? And as soon as this is done to free the memory?

If you can give some hints that would be awesome.