Relay vs Relax with the use case of whisper speech to text model for android phones

Questions
#1

I am trying to learn TVM with a task of porting the whisper speech to text model to android phone. I understand there are two ways to address my problem i.e relay → TIR and then relax → TIR. Which approach should I consider? If it is relax, can somebody suggests some examples. I tried to understand mlc-llm but still it is not straight forward.

1 Like
#2

if you are working on foundational models that involves dynamic shape(e.g. whisper) likely relax is something that would be relevant. We also have an in progress update of mlc-llm through nn.Module that hopefully makes some of the porting easier

1 Like
#3

Hi, was there any progress on whisper using tvm? I’d like to contribute. Thanks!