Open sc-aharri opened 4 days ago
Hi @sc-aharri, if you are able to staticize as much as you can, that'll probably be the easiest work around for now. I was able to reproduce on nightly with the exact same code... it might be at root a torch_xla issue, assuming we are passing in the right values here: torch_xla_utils.py
Description of the bug:
When I convert a pytorch model containing a MaxPool2D module,
ai_edge_torch.convert
crashes. This can be reproduced on my setup using the following minimal repro:The relevant stack trace can be found below.
Actual vs expected behavior:
The script crashes with the following callstack:
It seems that
is_unbounded_dynamic
is returning false. There are cetrtainly dynamic dimensions, but I'm suspicious of the word unbounded, since I am defining bounds. Should I expect bounds to be propagated to torch_xla?Any other information you'd like to share?
My setup is an Ubuntu docker container with cpu-only versions of torch.