cc @gante for generation config as well

cc @LysandreJik as well!

Most of the failing tests seem unrelated, will fix the unprotected import.

@Manalelaidouni added a bunch of comments with more precise limits :)

LMK if you need help getting the CI green ✅

When working on the improtected import, I implemented the lazy code logic for Pydantic in the main init and the generation module’s init, however the dummy routing only works when importing from transformers top level and not the generation module’s level, because there are other non-pydantic dependent objects that directly import from configuration_utils.py (e.g, NEED_SETUP_CACHE_CLASSES_MAPPING, QUANT_BACKEND_CLASSES_MAPPING, GenerationMode, etc) and bypass the lazy loading so configuration_utils.py is always executed.

This means I either need to move GenerationConfig to a seperate module so it won’t force a direct import, or guard GenerationConfig inside if is_pydantic_available and make all GenerationConfig imports from the main init instead of relative imports :

if is_pydantic_available():
     from pydantic import BaseModel, ConfigDict, PrivateAttr, confloat, conint, model_validator
	  class GenerationConfig(BaseModel, PushToHubMixin):
	       ...

Although, this means the nice import error msg in requires_backend won’t be triggered because the dummy GenerationConfig won’t be actually instantiated.

Thank you guys and let me know what you think before I proceed :)

@Manalelaidouni since the GenerationConfig object will now depend entirely on pydantic, I think we can promote pydantic to a mandatory requirement and get rid of the import guards. Let me confirm internally.

@Manalelaidouni We (me, @LysandreJik, and @ArthurZucker) discussed what to do regarding install requirements internally. Adding a mandatory requirement to a popular library like transformers is not a decision we should make lightly; it may place unwanted constraints on many projects.

That being said, we want to add more validation to certain classes (GenerationConfig, PretrainedConfig, processors, ...), and pydantic seems like the right tool for it. If we don't make it a mandatory requirement, we'll have to add an import guard in most files in our library, which is also unwanted.

Here's what we're going to do: I'm going to open a GH issue (and share it on social media) to see what the community thinks and to try to anticipate problems. Then, in about a week, let's make an informed decision 🤗

Hey @Manalelaidouni 👋

Thank you for holding this discussion with us -- we've decided no NOT move forward with pydantic. See my comment here 🤗

I opened huggingface/huggingface_hub#2895 as a suggestion on how to move forward with data validation in the HF ecosystem. Open to feedback and subject to change based on community requirements 🤗

Closing this one since the new solution is being implemented.

hi
my transformers is 4.57.3. when i using

 model = Qwen2ForCausalLM.from_pretrained(
            llm_path,
            config=llm_config,
            torch_dtype=torch.bfloat16,
        )

I encountered the problem:

   `cannot import name 'NEED_SETUP_CACHE_CLASSES_MAPPING' from 'transformers.generation.configuration_utils'`

how can i solve this