[Llama4] Remove image_sizes arg and deprecate vision_feature_layer

[yaswanth19]

As per the titles, simply removes image_size arg from Llama4's forward method as it's unused and deprecates vision_feature_layer arg.

@ArthurZucker for review

[yaswanth19]

vision_feature_layer is also not used in get_image_features and thus not necessary. Should I remove it, since we have kwargs here which should consume that whenever passed thus not breaking BC.

[ArthurZucker]

yep sounds good IMO 😉

[Rocketknight1]

cc @zucchini-nlp as well

[zucchini-nlp]

Looks like a bad copy from Llava, and can be simply removed. I don't expect anyone to be using vision_feature_layer, so let's just remove it since it is inn utility method

[yaswanthg15]

@zucchini-nlp It's also an arg in forward and present in config as well, shall I remove from forward too, removing from config would be breaking. Otherwise def can only remove from utility func.

[zucchini-nlp]

It's also an arg in forward and present in config as well, shall I remove from forward too, removing from config would be breaking.

Ahh I see, then we can do a tiny deprecation cycle of 1 minor release. Prob ir was also serialized in config in which case it will be still available with config.vision_layer. So we just need to raise warning when users try to set or get unused attributes

[ArthurZucker]

Thanks, not sure about the warning because if we pushed this to the hub, it is probably something users can do nothing about. We don't need a warning as the arg will always be read in **kwargs

[yaswanthg15]

@ArthurZucker Then you are suggesting that, we remove the arg from config completely (will be read in kwargs for all hub configs) and add the deprecate decorator in modeling file?

[zucchini-nlp]

i think we just need warning when users try to get the attribute or set to a different value, letting them know that vision_feature_layer has no effect on model

[yaswanth19]

@zucchini-nlp A code suggestion or reference would be better here. I can only think of a niave way to have getter and setter funcs for vision_feature_layer attr and raise a warning like you mentioned (instead of general warning) or is there a util func or decorator which can be used ootb

[zucchini-nlp]

Smth like this

transformers/src/transformers/models/gemma3/configuration_gemma3.py

Lines 240 to 251 in 3fd456b

	@property
	def sliding_window_pattern(self):
	warnings.warn(
	"The `sliding_window_pattern` attribute is deprecated and will be removed in v4.55.0.",
	FutureWarning,
	)
	return self._sliding_window_pattern
	@sliding_window_pattern.setter
	def sliding_window_pattern(self, value):
	self._sliding_window_pattern = value

[zucchini-nlp]

Thanks, the generation test will be fixed soon and is not related :)

[zucchini-nlp]

Lets rebase and merge :)

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: llama4

[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.