make all-linear as default for target_modules

[pacman100]

What does this PR do?

1. Make all-linear as default for target_modules. This should make LoRA and variants work for any new model out-of-the-box.

[HuggingFaceDocBuilderDev]

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.

[BenjaminBossan]

Hmm, I'm not sure, this would break backwards compatibility, right? If we have a supported model architecture that does not target all linear layers but only a subset, using the LoraConfig defaults would suddenly target different layers. I think this would be very unexpected for users.

We could consider to leave None as default but to fall back on all-linear in case that the model architecture is not supported. But even then there can be problems. For instance, all-linear currently only works for transformers models. Also, as we saw in #1485, the function to detect all linear layers is still not quite robust. I thus don't think it's mature enough to be used as default.

My personal preference would be to leave the default the same. We could, however, improve the error message in case that the architecture is not supported by adding something like "When using a transformers model, consider setting target_layers="all-linear" to automatically add LoRA to all Linear layers".

[pacman100]

Thank you Benjamin, makes sense, closing the PR.