AOT is very slow

#1

I found that AOT compiler code is very very slow. I took more than 3 hours to compiler tensorflow inceptionV3 model.
So there any method to reduce the compiler time?
@MarisaKirisame

#2

are you calling the compiler on device? It is indeed slow, and another option is to use the relay virtual machine instead.

#3

Thank you very much,

#4

It would be nice to have an AoT compiler that supports cross-compilation. Is that currently in the AoT compiler timeline @MarisaKirisame , or will that be deprecated by the Relay VM?

#5

I am very busy for 3 months - as you can see, I am not even updating my PR.
It mean aot wont receive any >1day update for 3 months from me.
Personally, I think it is getting deprecated by the VM - it will still be there and usable, but as long as VM isnt significantly lower, there is no reason to stay.

#6

@MarisaKirisame it makes sense to invest more time into the VM, and plan on deprecating the AoT compiler moving forward.