Fix: generate() with max_new_tokens=0 produces a single token.

[danielkorat]

What does this PR do?

Currently, setting max_new_tokens=0 produces 1 token instead of 0, and the warning is unclear.
For example, for the following code:

checkpoint = "bigcode/tiny_starcoder_py"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint)
inputs = tokenizer("def print_hello_world():", return_tensors="pt")
max_new_tokens = 0
outputs = model.generate(**inputs,
                        pad_token_id=tokenizer.eos_token_id,
                        max_new_tokens=max_new_tokens)
input_length = len(inputs['input_ids'][0])
output_length = len(outputs[0])
print(f"\nTest:{input_length - output_length == max_new_tokens}")

The output is:

utils.py:1134: UserWarning: Input length of input_ids is 7, but `max_length` is set to 7. This can lead to unexpected behavior. You should consider increasing `max_new_tokens`.
  warnings.warn(

Test: False

After the fix, this is the output:

`max_new_tokens`=0, no tokens will be generated.
utils.py:1134: UserWarning: Input length of input_ids is 7, but `max_length` is set to 7. This can lead to unexpected behavior. You should consider increasing `max_new_tokens`.
  warnings.warn(

Test: True

(Note the new warning).

Currently fixed only for greedy_search(). Once this PR is reviewed, I'll add the fix to all other generation modes.
@gante @amyeroberts

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.

[gante]

Hi @danielkorat 👋

In lower-level APIs like generate, we should strive to be more strict if some parameterization is causing problems to avoid creating unexpected patterns. In this case, generating with 0 tokens should not be possible, as there is no generation involved.

As such, I'd suggest increasing the severity of the associated warning to an exception instead of the changes proposed here :)

[danielkorat]

Hi @danielkorat 👋

In lower-level APIs like generate, we should strive to be more strict if some parameterization is causing problems to avoid creating unexpected patterns. In this case, generating with 0 tokens should not be possible, as there is no generation involved.

As such, I'd suggest increasing the severity of the associated warning to an exception instead of the changes proposed here :)

Hi @gante, I made the requested change in a new PR: 28621.