adding whole Linear8bitLt/Linear4bit module save/load serialization

[rdyro]

The purpose of this pull request is to allow torch.save/torch.load directly on modules containing Linear4bit and Linear8bitLt submodules.

Currently, torch.save, then torch.load on Linear8bitLt (after first forward) causes a missing field CB error in the Int8Params class. This PR makes torch aware of the CB and SCB fields in Int8Params class.

The core of this PR is

        ~~return torch.Tensor._make_subclass(cls, data, requires_grad)~~
        obj = torch.Tensor._make_subclass(cls, data, requires_grad)
        obj.CB, obj.SCB = cls.CB, cls.SCB
        return obj

in class Int8Params

I also added the torch.save -> torch.load test to the Linear4bit (this was already working) and Linear8bitLt (this is not yet working).

While saving modules directly in Pytorch with save and load is not good practice, the change to make this work is minimal and makes disk caching modules for development easier.

[younesbelkada]

cc @Titus-von-Koeller wdyt? might be good to have for the next release no?

[Titus-von-Koeller]

Yes, I agree, this is looking good and should be merged before the release. I'll review it more in depth soon.

Thanks @rdyro for the good work and taking the initiative to contribute, really appreciated 🤗

[rdyro]

Thanks for the positive feedback! I really like your work with bitsandbytes.

Let me know if ideally, the new tests should extend to all Linear layers, not just Linear4bit and Linear8bitLt.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Titus-von-Koeller]

Dear @rdyro,

I just reviewed your proposed changes and everything really looks good! I don't think any additional tests are needed, what you did already looks good the way it is.

I also ran the transformers integration tests and everything came through clean.

Thanks so much for your contribution and if you feel like contributing more, we'd be happy to support you!