Fix torchscript errors

[marksgraham]

I've been getting torchscript errors when runnng tests on my machine that @Warvito does not find, as documented here. It would be good to get to the bottom of why these errors are appearing on some setups and not others.

[Warvito]

I started to look how github actions is configured in MONAi core, so we can deploy it here too and have the tests running in the same environment (related to #185).

As Eric said, we can try the MONAI docker image to run the tests locally too.

[ericspod]

Getting tests up and running would be great to track down these issues. I found one with global variables which can't be packaged in Torchscript objects, so has_xformers here will cause the model to not be compatible with torch.jit.script. I would suggest that the choice of using one component or another should be defined at the global scope or global values like this stored as attribute variables of the object.

[Warvito]

Getting tests up and running would be great to track down these issues. I found one with global variables which can't be packaged in Torchscript objects, so has_xformers here will cause the model to not be compatible with torch.jit.script. I would suggest that the choice of using one component or another should be defined at the global scope or global values like this stored as attribute variables of the object.

@ericspod Does allowing the user to define if he wants to use memory efficient attention when creating the object and checking in the init() method if the dependency exists sounds good (as @danieltudosiu sugested here)?

Not sure if I configured something wrong. Should the min tests raise an error in this case?

[ericspod]

@ericspod Does allowing the user to define if he wants to use memory efficient attention when creating the object and checking in the init() method if the dependency exists sounds good (as @danieltudosiu sugested here)?

Not sure if I configured something wrong. Should the min tests raise an error in this case?

I think it would be nice to provide the user with a choice somehow just in case.

I can't replicate the issue locally by copy-pasting the tests using test_script_save so it's possibly an issue with the Pytorch version. I need to check another machine what version of Pytorch I'm using there to see if it reoccurs.

[marksgraham]

I've checkd and am no longer getting these errors. I'm going to close for now unless this issue crops up again (unless you guys want to continue discussing other torchscript issues here?)