Fallback of jit compilation

[MBPhys]

Hi,

first of all, congratulation to this amazing project 😄
Unfortunately the jit compilation doesn't work. There is a fallback.

It seems that the Bound_class is the problem, because it is not a tensor directly.

Do you have an idea how we can fix them?

Best regards

Marc

[balbasty]

Hi Marc

Thanks for reporting this. Could you please tell me which version of PyTorch (and which platform) you are using?

(edit: and how you call the function: inputs, etc)

Thanks
Yael

[balbasty]

This issue seems to be related to pytorch/pytorch#82889
It only happens in recent versions of pytorch (I reproduced it on 1.12, but I don't know when it was introduced).

In the meantime, the suggested workaround (set the environment variable PYTORCH_JIT_USE_NNC_NOT_NVFUSER=1) seems to work.
I could set it from inside torch-interpol (on import) but I am not sure it's the best solution.

[ZacharyVarley]

Too add, I also was having some issues with NVFUSER but torch 2.0.X seems to have fixed whatever the issue was.