Deprecate #36741 and map Causal to Conditional

[zucchini-nlp]

What does this PR do?

Fixes #36886, fixes #36926 and loading in SmolAgents (from feedback in internal slack)

NOTE: this is a temp fix. In long term we will converge under one Auto for all multimodals, which we don't have yet. As such we will keep CausalLM as temporal dump for custom code users. 🔴 We will break this pattern in the near future and might enforce text-only models under this mapping

After #36741, we unintentionally broke model loading for most remote code users, because many Vision/Audio/Omni LLMs on the hub use CausalLM mapping with AutoTokenizer. This happens because AutoImageTextToText is less visible, and also because we have no mapping for other modalities.

This PR deprecates the previous fix and properly maps Gemma3 4B+ models to its ConditionalGeneration class which aligns with info in model card. As discussed internally, all vision-audio-multimodal models will converge under AutoModelForCausalLM in the future to maintain consistency and stop adding new mapping for each new modality.

• Why? All Audio LMs are already in causal mapping due to the lack of AutoAudioTextToText or AutoAudioToText mappings. Vision LMs have also been inconsistently mapped under CausalLM

• Consequences: No breaking changes for users, including those using remote code. A warning is raised only if a model has both full config and text config mapped in CausalLM which is a super rare case. (gemma-3 was exception)

• Edge cases: Other models like llava-1.5 cannot load anymore under AutoModelForCausalLM after this PR, but checkpoint keys won’t match anyway + we never got user issues. This was an existing issue, not introduced by this PR. It might be resolved by a bigger refactor for vLLM after adding base models for all VLMs and correct base-prefix-keys

I verified that Gemma3 case works as expected, without raising warnings (I just changed mapping class). The only difference from the previous fix is that vision tower is loaded as well, which might affect advanced users who manipulated configs or access model layers manually.

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained('googel/gemma-3-4b-it')
tokenizer = AutoTokenizer.from_pretrained('googel/gemma-3-4b-it')
inputs = tokenizer('What is your name?', return_tensors="pt")
out = model.generate(**inputs)
tokenizer.batch_decode(out)
>>> ['<bos>What is your name?\n\nI am called Pixel.\n\nHow are you?\n\nI am functioning optimally, thank you for']

# Chat template
msg = [{"role": "user", "content": "What is your name?"}]
tokenizer.apply_chat_template([msg], tokenize=True, return_dict=True, return_tensors="pt", add_generation_prompt=True)
out = model.generate(**inputs)
tokenizer.batch_decode(out)
>>> ['<bos><start_of_turn>user\nWhat is your name?<end_of_turn>\n<start_of_turn>model\nMy name is Gemma. I was created by the Gemma team at Google DeepMind. \n\nI']

[github-actions]

Hi 👋, thank you for opening this pull request! The pull request is converted to draft by default. When it is ready for review, please click the Ready for review button (at the bottom of the PR page).

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ArthurZucker]

TLDR of our internal thread:

• we need a AutoForAny, as today AutoForCausalLM is a dumpster used to map anything
• we cannot break, so for now it will remain this way
• we also want text only parts to be loadable only, similarly you would want to load only Image and Text -> ImageTextToText.

Let's patch this!