Hi,
We’re just starting to look into Relax for e2e model compilation and I’ve some basic questions. I would greatly appreciate if you could share some insights on the these:
-
AFAIK, when we translate Relay into Relax, it uses OpStrategy to find the OP implementations, however, it appears to use only compute and ignore schedule. This means that if we want to make any scheduling changes to the generated STIR for relax, it can be done only via a TIR pass or metascheduler/auto-tuner. Could you please confirm if this understanding is correct?
-
Can we use Relax to compile a model e2e without using MetaScheduler and execute it on VM? I’ve found some examples in the tests/python/relax using VM (without using auto-tuning) but these’re very small, single op examples. Just wondering if someone has tried using this workflow for a full model and can share their observations
Thank you so much! -Jyotsna