Attn implementation for composite models

[zucchini-nlp]

What does this PR do?

Related to #32221 and fixes #30565. As described in the issue, currently MultiModal models set attn implementation only in the LM backbone, and not in vision backbone. While working on it, I found another bug. Specifically, using a property to set SDPA flag to VLM class doesn't work because we don't know what is self.LM before init the VLM. I tried class property, but that would throw error that self has no attribute LM.

This PR makes a few modifications to how attn impl is set in modeling to fix the above issues.

More precisely:

• If the model is composite, i.e. has 2 or more PreTrainedConfigs as part of the main config (text/vision config), then we set the requested attention implementation to each of the sub-configs as a dict. For ex, in LLaVa case:

model.from_pretrained(id, attn_implementation="sdpa")

# internally we do the following, where the second line dispatches SDPA on LM backbone
# if we can't dispatch, an error is raised as it should be when loading an LLM with unsupported attention (i.e. llama doesn't support SDPA)
self.config._attn_implementation = {"text_config": "sdpa", "vision_config": "sdpa"}
AutoModelForCausalLM._from_config(config.text_config, attn_implementation=config._attn_implementation["text_config"])

• If user just loads the composite model without any explicit attention, then we dispatch SDPA in any of the sub-models whenever possible. In case any/none of the vision/text models support SDPA, we silently fallback to eager. No warnings, no errors, which is in line with non-composite models

# suppose LM backbone doesn't support SDPA here, then we'll use eager silently
# but dispatch SDPA on vision backbone
model.from_pretrained(id, attn_implementation=None)

• The users are also free now to set different attentions in different backbones. We can set eager on vision model and sdpa on LM. That can be done if one passes a dict attn_implementation. And the dict keys have to be identical to the ModelConfig attribute names for sub-configs. In other words, if the config has config.text_config, then the dict key must be "text_config" to correctly dispatch

model.from_pretrained(id,  {"text_config": "sdpa", "vision_config": "eager"})

• For non-composite models nothing changes and everything works as before

NOTE:
I had a few rare cases where the model was not really composite but had several sub-configs. For example DBRX, and I decided to get rid of the sub-configs there. I am not sure what is the right way to deprecate that, so any comments on that welcome
Also, some models have nested configs, where a vision config is part of text config (Qwen2-VL). Only three models are like that, mostly because I didn;t review properly when the model was added and didn't insist on changing stuff. For those models, we don't consider them as composite and just leave it working as it was

[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.

[qubvel]

Thanks for working on this! Not sure I fully understand the logic at this moment, so I left a few comments below

[ydshieh]

Regarding

MODEL_MAPPING.get(type(config.text_config), None)
            if text_model_cls is not None:

it could probably reuse

def _get_model_class(config, model_mapping):

But both will be somehow slow as you mentioned, and sometimes give unexpected errors (because loading a modeling module file). But we might have someway to avoid that.

[zucchini-nlp]

@ydshieh @qubvel I made some changes after our last discussion that changing class attr is not good. So now we don't change the attr per se, but we change the config's attn_impl manually while setting sdpa. I checked that it works this way also. The last question is whether to set it to SDPA when both sub-configs support it or at least one supports it.

I will add more tests if we agree on the design

And also, regarding

But both will be somehow slow as you mentioned, and sometimes give unexpected errors (because loading a modeling module file). But we might have someway to avoid that.

unfortunately we can't do that, cause it raises error if key is not in the dict and we have not auto-mappable models. So we rather try to get and if not simple continue.

[ydshieh]

unfortunately we can't do that, cause it raises error if key is not in the dict and we have not auto-mappable models. So we rather try to get and if not simple continue.

but why not try except around the usage of _get_model_class? I don't feel very strong here, but just it is something well designed in order to get the model class which is good if we can reuse it.

[zucchini-nlp]

Hmm, not sure how good if to use a try except, can it be another source of hidden bugs even if we use except AttributeError? I'm okay with any way, as long as it's consistent with the library and doesn't cause bugs

[zucchini-nlp]

As we discussed internally, I made the loading logic model-agnostic. So now we check if the config contains any PretrainedConfig inside and set attn implementation to each PretrainedConfig. This solved vision-encoder-decoder problem we had in slack and makes us more versatile for possible future multimodal models.

I added some tests in vision enc-dec models where only one sub-config supports SDPA and when both support. Also added tests in Idefics (which were skipped earlier) to verify that LM has its SDPA dispatched even though other modules do not.

TODO in the next few hours or tomorrow:

• Add similar tests to encoder-decoder models, not only vision enc-dec
• Accept from users a config with separate attn_implementation already set, such that attn_implementation={"vision_config": "sdpa", "text_config": "eager"}

[zucchini-nlp]

@ArthurZucker I made changes as you suggested but I want to warn that it still needed a few extra logic. Making this PR revealed soo many issues. So we have two options now to indicate attn for composite models:

Attn implementation as a dict:
The prev solution when we simply copy paste the user-requested attn in all sub-configs, and form a dict. So that the general config.attn = Dict(text_config: None, vision_config: None)

• no need for flag whether attn implementation is set or not, because when we have a dict we always use the very first requested attn in all sub-configs. We just need to save it as a dict once and then go back to that dict to "get()"
• we don't need to unset attn implementation when saving so one can set a different attn when loading any model back.

• BUT we change a lot of model code to pass in explictly (attn_implementation=config._attn_implementation["text_config"]) and we need to account for cases like Chameleon or DBRX where the sub-configs don't even need attn

Attn implementation directly propagated in corresponding configs:
Current solution when we simply propagate the attn to all configs we can find, and don't touch the general config

• no need to change any modeling code, except for general clean up

• BUT we need a flag attn_implementation_autoset so we don't dispatch attn on the same config twice. Also we would need to unset the flag when saving. See below for reason

Imagine we have a composite or nested config. Currently we set attn in three places: from_pretrained, from_config and init. When we don't require any attn (None) and load with model.from_pretrained(id), internally we will call autoset_attn and set SDPA to the general config and the sub-configs will get None propagated. But then the same method autoset_attn is called second time when we init the class, and this time we'll try to propagate SDPA to all sub-configs (because general config has SDPA already set). If any sub-module has no SDPA support, an error will be raised as it was requested directly. From user perspective they did nothing wrong, but they see an error.
So the decision is to either stop autosetting attn for each config once it was set, which is happening now. Or in prev version I had attn implementations saved as dict, which basically did the same. When we got a dict attn, we had nothing to set and simply returned

[ArthurZucker]

Modeling changes are just perfect 😉 before we have a big refactor this is super super welcome

[ArthurZucker]

If this is donc to set the attn_implementation of the sub configs, I am wondering if we can't have a way to use is_composition to simply say:

• if we are a composition, then we call on our composiite member the _autoset_attn_implementation function

Today we have a problem with is_compoisition: we have this

    @classmethod
    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
        cls._set_token_in_kwargs(kwargs)

        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)

        if config_dict.get("model_type") == "mllama":
            config_dict = config_dict["vision_config"]

        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
            logger.warning(
                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
            )

        return cls.from_dict(config_dict, **kwargs)

which IMO we should never have to do and we copy past this everywhere.

This means there is something wrong with how we handle composition!

[ArthurZucker]

If we find a clean way to auto_set attn implementation would be better in the futur!

[zucchini-nlp]

merging tomorrow, the tests pass locally in most cases except fro the cases not touched by this PR. For ex some FA2/sdpa equivalence tests are failing for me in main branch