Fixing question-answering with long contexts

[Narsil]

Fixes #13811

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[sgugger]

Thanks for fixing, nice new tests!

[sgugger]

Why shallow for the name?

[Narsil]

Could be a better name, I just tried to disassemble the feature before passing it on for _forward as it makes it benefit automatically from no_grad and to(device) from forward (otherwise _forward had to take care of it).

Reusing features is good enough? Didn't find a better name.

[sgugger]

I like features personally (and I use example/sample when I want to make the distinction: one sample gives several features).

[vikramtharakan]

Having a similar issue where I used a QA model from transformers, tweaked the model in label studio making some annotations and then tried to load the model back again. The model is pulling out the correct answers but seemingly hadnle_impossible_answers isn't working because it gives an answer for every questions even when the question is irrelevant.. What's even weirder is that it doesn't do this in label studio's interface so seemingly this handle_impossible_answers is working on that side. My contexts I pass through the pipeline are also quite long

Made sure I have the most up to date transformers model, and the models were saved out with save_pretrained

model_path = "PATH"
model = AutoModelForQuestionAnswering.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path) 
QA = pipeline('question-answering', model=model, tokenizer=tokenizer)

# Other code...
answers = QA(question=questions, context=context, top_k=1, max_answer_len=32, handle_impossible_answer=True)

The answers that I get out of here have an answer for every question, even when it doesn't need to be answered. I've tried loading the models multiple ways - the only way I got handle_impossible_answers to work as intended is if the tokenizer wasn't the AutoTokenizer that I had used here. But then the answers it gave me were complete garbage. Anybody else run into this issue with AutoTokenizer??

[Narsil]

The only was I can think the tokenizer could be involved would be if it doesn't include EOS/BOS.

For null answer to come out, the score has to be the highest when the start and end logits are the first one (which should be a BOS/CLS token). Did you finetune your model on such null answers ? Maybe the finetuning just made that output impossible for the model to produce ?