Are XLA AOT-compiled models self-sufficient or do they still require a TF runtime?

TensorFlow General Discussion
1

Hi,

I am not familiar with XLA AOT so please excuse my ignorance, however I went through the documentation/tutorial but couldn’t find the information.
The question is:
Are XLA AOT-compiled models self-contained? I mean, do they need an extra runtime? If not, do they only contain the relevant operation code or also extra code (other ops, device management, etc).

Thank you in advance

3

I read the tutorial again and it says clearly that the models only contains kernel code.