[GenerationConfig] add additional kwargs handling

[ArthurZucker]

What does this PR do?

This add the same support that we have in the PretrainedConfig, where additional kwargs are automaticallu updated.
This will allow users to re-use the GenerationConfig class for most of the use_cases, whithout having to add a model specific class. I was trying to load the following generation_config and got half of my additional arguments deleted 😉

[gante]

We probably can remove self.generation_kwargs = kwargs.pop("generation_kwargs", {}) on L277, it was intended as a restricted version on these changes

(for context, the model config has the same lines here)

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks for adding this!

[ArthurZucker]

Also will have to add test + this is apparently breaking a lot of things haha

[ArthurZucker]

Okay, after talking a bit with @gante and testing, this is not the best, this PR will focus on other missing functionalities. Mostly addition of the dict_torch_dtype_to_str function, as the dtype could be passed to the generation 😉
The problem is mostly that if we process all the additional kwargs, we are getting all of the arguments from the configuration.json which mixes things up.
The simplest solution is either to store them in generate_kwargs or re-write the configuration for the model. I though this was cumbersome but it is actually the most logical and cleanest way to do it.

EDIT : gonna just add a condition, if the kwargs are from a config file, they are not added.

[ArthurZucker]

Now only thing left is to add a pretty test with all the different edge cases I encountered.

[gante]

This line now only exists to obtain unused_kwargs, as the kwargs get written to the config in the line above, correct?

[ArthurZucker]

Well yes, for example when _from_model_config is set to True, then it is still in the kwargs, think I saw something like this.

[gante]

I'd add a comment here explaining why we need this if, otherwise we may be like "wtf?" in the future

[ArthurZucker]

Sure, thanks for the comment!