🎩 ⛷️ Switch to using inference_mode instead of torch no_grad

[sbonner0]

Description of the Change

Replaced the use of torch.no_grad with the newer torch.inference_mode call. More details on inference mode can be found here: https://pytorch.org/docs/stable/generated/torch.inference_mode.html

Essentially it is an even more extreme no_grad method and can give free speedup to inference tasks. Hopefully future tweaks for this mode will come from the pytorch team so would be cool to switch to it.

Possible Drawbacks

I think this was only introduced in pytorch 1.9 so would limit what versions of pytorch can be used with pykeen. It's also possible that some of the limitations of inference mode -- for example "tensors created in inference mode will not be able to be used in computations to be recorded by autograd after exiting inference mode." may have some unintended consequences down the line.

Verification Process

Ran a few training and evaluation loops with existing code I had. Predictive performance results remained the same...also didn't see a massive amount of speedup for inference using a 1080Ti.

[mberr]

Sounds good for me. From my point of view, introducing a torch>=1.9.0 requirement is not a problem - we generally develop code to be compliant with the latest stable release of the dependencies although, of course, it is nice if they also work with older versions.

[mberr]

this requirement is already in the setup.cfg , cf.

pykeen/setup.cfg

Line 61 in 2e12cfc

torch>=1.9; platform_system != "Windows"

[cthoyt]

@PyKEEN-bot test

[cthoyt]

@PyKEEN-bot test

[cthoyt]

Thank you @sbonner0!