[RAG] Fix various generator issues

[lalitpagaria]

Resolve #587 and #585

for num_return_sequences refer.
https://github.com/huggingface/transformers/blob/a1bbcf3f6c20e15fe799a8659d6b7bd36fdf11ed/src/transformers/modeling_rag.py#L852
This should not be greater than num_beams hence when use pass top_k value greater than num_beams we reset it to num_beams and print warning.

[tholor]

The top_k fix seems good and straightforward.

I have some doubts about the GPU fix though. With removing the .to(self.device) calls, the tensors should still reside on the CPU. Haven't tested this, but I suppose this solves the error msg in #587 by not utilizing the GPU at all anymore. We should instead make sure that all required tensors are on the GPU (= adding more to(self.device) calls). Let me know if I should jump in and help to code this up.

[lalitpagaria]

Yes please. I tried to debug but couldn't as I don't have GPU. But I would suggest to create another PR to enable GPU support for generator

[tholor]

Ok then let's keep this one limited to the top_k fix (and remove the GPU changes). I will raise a new one for GPU support and fixing #587.

[lalitpagaria]

I would suggest to keep these changes as well because they will allow people to use generator on CPU and GPU without error (even with low speed).

[tholor]

Well, isn't there only an error if the user set use_gpu=True (on a GPU machine)? I think it's better to throw an error in that case than "failing silently" and not using the GPU at all. Otherwise, people wonder why it is so slow.

[lalitpagaria]

Agree. I will remove GPU related changes and raise exception if user try generator with use_gpu=True

[lalitpagaria]

Now exception will be raised if user use use_gpu=True for RAGGenerator

[tholor]

Looking good!