Incorrect Pytorch version check while loading model

[RIP-Skillan]

In whisper/__init__.py line 150, for adding the 'weights_only' argument the if condition checks the python version using a simple string comparison as shown below:

with (
        io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")
    ) as fp:
        kwargs = {"weights_only": True} if torch.__version__ >= "1.13" else {}
        checkpoint = torch.load(fp, map_location=device, **kwargs)
    del checkpoint_file

Since string comparison uses lexicographical ordering the comparison returns incorrect results.
For example: 1.8.1 > 1.13.0, which is wrong since by PEP standards 1.13 comes after 1.8.

This can be fixed by modifying the code as follows:

from pkg_resources import packaging
with (
        io.BytesIO(checkpoint_file) if in_memory else open(checkpoint_file, "rb")
    ) as fp:
        kwargs = {"weights_only": True} if packaging.version.parse(torch.__version__ ) >= packaging.version.parse("1.13") else {}
        checkpoint = torch.load(fp, map_location=device, **kwargs)
    del checkpoint_file

I would like to create a PR for the same. Would this be a good solution or is there any better ones?
I've never contributed before, so I would like as many suggestions as I could get

[shiv669]

You’re right — comparing version strings directly is unsafe because of lexicographic ordering (e.g. "1.8.1" > "1.13.0"). The standard way to handle this is to parse versions properly. Your solution using packaging.version.parse is correct and aligns with PEP 440.

from packaging import version
kwargs = {"weights_only": True} if version.parse(torch.__version__) >= version.parse("1.13") else {}

That’s a clean fix and definitely PR-worthy. Just make sure to import packaging.version (not pkg_resources.packaging) since packaging is the modern maintained package.

[RIP-Skillan]

Oh okay, thank you for helping me out.
I'll make the changes.