Checkpointer load should specify `weights_only`

[baldassarreFe]

From pytorch 2.6 (next release), torch.load will switch the default for weights_only from False to True. I was training models using 2.5 (which defaults to weights_only=True) and when I switched to 2.6 I got an exception when resuming a checkpoint because fvcore doesn't pass that parameter and the the default is now False:

fvcore/fvcore/common/checkpoint.py

Line 254 in a491d5b

return torch.load(cast(IO[bytes], file), map_location=torch.device("cpu"))

I think Checkpointer should allow passing a value for weights_only, either in the constructor or in Checkpointer.load().
It's a trivial modification to the code, but it's worth discussing where to introduce the new parameter before proposing a PR.

[baldassarreFe]

Any update? What is the recommendation?

• add a weights_only parameter to the constructor
• add it to the load() function
• don't add it and close the issue, users can just subclass the checkpointer and modify the behavior themselves