Widen match condition for _can_record_outputs

[molbap]

In case of module reloading, we currently lose tracking hooks for hidden states and attentions . Widening the matching condition a bit.

Should fix #43761, also mentioned in linkedin/Liger-Kernel#1061

Reproducer for CLIP:

import importlib, torch
from transformers import CLIPVisionConfig, CLIPVisionModel
import transformers.models.clip.modeling_clip as modeling_clip

config = CLIPVisionConfig(hidden_size=32, intermediate_size=64, num_hidden_layers=2, num_attention_heads=4, image_size=30, patch_size=10)
pixel_values = torch.randn(1, 3, 30, 30)

m1 = CLIPVisionModel(config)
o1 = m1(pixel_values=pixel_values, output_hidden_states=True)
print("before reload:", o1.hidden_states is None)

importlib.reload(modeling_clip)

m2 = CLIPVisionModel(config)
o2 = m2(pixel_values=pixel_values, output_hidden_states=True)
print("after reload:", o2.hidden_states is None)

Second reload will currently fail, and this is a fix proposal.

Added a test as well (specific to CLIP, likely enough)

[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.

[zucchini-nlp]

This line of check for class_name was added for AyaVision, but I think it's not needed anymore. At least tests don't complain 😄

[molbap]

aah :D makes sense, wdyt of the added check though? not sure how to handle reloading more elegantly

[zucchini-nlp]

I think it's the best we can do with importlib.reload and shouldn't backfire with false positives. Lemme approve, forgot to

[zucchini-nlp]

maybe we can move it under test_modeling_common.py, for make smth similar to this test and reload the module at some point before checking if config.output_attentions works?

transformers/tests/test_modeling_common.py

Lines 1754 to 1757 in ace7c37

	def test_attention_outputs(self):
	if not self.has_attentions:
	self.skipTest(reason="Model does not output attentions")

[molbap]

yeah I hesitated, I didn't want to add yet another test for 400+ models 🫣 but you may be right

[zucchini-nlp]

haha right!

[ArthurZucker]

great catch actually! 🤗

[molbap]

I want to think that great minds think alike 👀 @Cyrilvallez 's PR solves more problems with check_model_inputs #43765