To unify text_generation()'s stop_sequences and TextGenerationInput's stop params

[sadra-barikbin]

Hi there! 🤗

It would be great if text_generation()'s stop_sequences and TextGenerationInput's stop params are unified just like chat_completion and ChatCompletionInput to be able to feed the input dataclasses to the client methods directly:

client.text_generation(prepared_request.inputs, **prepared_request.parameters) # currently raises `TypeError: text_generation() got an unexpected keyword argument 'stop'`
# or
client.chat_completion(**prepared_request)

huggingface_hub/src/huggingface_hub/inference/_client.py

Line 1658 in 359093f

stop_sequences: Optional[List[str]] = None, # Same as `stop`

[sadra-barikbin]

Currently I should do this as the alternative:

from dataclasses import asdict

input_as_dict = asdict(input)
input_as_dict["parameters"]["stop_sequences"] = input_as_dict["parameters"]["stop"]
del input_as_dict["parameters"]["stop"]
client.text_generation(input.inputs, **input_as_dict["parameters"])

[Wauplin]

Hi @sadra-barikbin, yes I definitely agree this parameter should be unified. I opened a PR to fix it :) #2473

[sadra-barikbin]

Thanks @Wauplin ! 👍

[sadra-barikbin]

There's also a discrepancy between ChatCompletionInput's stream param which is optional defaulting to None and chat_completion()'s stream param which should be either True or False. Is this OK?

huggingface_hub/src/huggingface_hub/inference/_generated/types/chat_completion.py

Line 123 in 359093f

stream: Optional[bool] = None

[Wauplin]

Thanks for flagging this @sadra-barikbin. It's a slight discrepancy but it allows us to simplify the @overload logic in InferenceClient.chat_completion. If we add the type annotation for None value, we would have to overload the method even more. The problem is that bool and Literal[True, False] is not considered the same by mypy / code editors. It led to quite some headaches when trying to solve that (the main thing I was focused on was to have correct autocomplete for the lib'). So all of this to say, the slight discrepancy is fine for now especially because it doesn't change anything in this case.