Allow bos_token_id is None during the generation with inputs_embeds

[LZHgrla]

What does this PR do?

Allow bos_token_id is None during the generation with inputs_embeds.

This is important for multi-modal inputs / generation for LLMs whose bos_token_id is None

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.

[amyeroberts]

Change looks good to me - thanks for adding!

Just need to add a test to make sure:

• Error is triggered when input_embeds is not passed and bos_token_id is None
• Not triggered when input_embeds is passed and bos_token_id is None

cc @gante to confirm the desired behaviour

[LZHgrla]

@amyeroberts @gante
Hi! I have added the related tests for this PR.

[gante]

Makes sense -- BOS is not required in all cases.

Thank you for fixing 🙏

[gante]

@amyeroberts should be ready for a final check

FYI @zucchini-nlp

[amyeroberts]

Thanks for working on this and adding a test!

Just a small comment on the test logic which I think has to be addressed before merge

[amyeroberts]

I think? Otherwise the error is being raised from the lack of inputs

Suggested change

	model.generate(max_length=20, bos_token_id=None)
	model.generate(input_ids, max_length=20, bos_token_id=None)

[LZHgrla]

@amyeroberts

I think we should not pass in input_ids. I tested the following code on the main branch and found that the generation works well (see case 3) If we pass in input_ids with bos_token_id=None.

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2")
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
article = "Today a dragon flew over Paris."
input_ids = tokenizer(article, return_tensors="pt").input_ids

# Case 1
ids = model.generate(max_length=20)[0]
print(tokenizer.decode(ids))
# �vedvedvedvedvedvedvedved Wh Wh Wh Wh Wh Wh Wh Wh Wh Wh Wh

# Case 2
ids = model.generate(input_ids, max_length=20)[0]
print(tokenizer.decode(ids))
# Today a dragon flew over Paris. fe fe fe fe

# Case 3
ids = model.generate(input_ids, max_length=20, bos_token_id=None)[0]
print(tokenizer.decode(ids))
# Today a dragon flew over Paris. fe fe fe fe

# Case 4, error
# Below code will raise a ValueError
# ids = model.generate(max_length=20, bos_token_id=None)[0]  

[amyeroberts]

Ah, you're right, sorry. Thanks for showing the cases so clearly!

[amyeroberts]

Thanks!