Fix/pe audio video bugs

[massimilianoviola]

What does this PR do?

1. Migrate PE-AV processor to the v5 sub-processor API

PeAudioVideoProcessor still uses the legacy feature_extractor_class and video_processor_class that #41633 deprecated, so every checkpoint load prints two deprecation warnings. The Auto mappings are already registered, so we can drop the legacy attrs and add an explicit __init__.

2. Fix get_*_embeds crash

The get_text_*_embeds helpers call the text model without output_hidden_states=True and then access text_outputs.hidden_states[-1], which is None, and so they crash with TypeError.
The fix is one extra kwarg per helper, mirroring the forward behaviour.

3. Friendlier forward error for single-modality inputs

forward requires ≥2 modalities; single modalities are handled by the get_*_embeds helpers.
Extended the ValueError to mention the existence of those helpers, so users reading the "you can omit any of the modalities, and use the same forward method" in the model card aren't stuck.

4. Fill in docs/source/en/model_doc/pe_audio_video.md

Replaced with a short overview, the clean architecture figure (upload pending here https://huggingface.co/datasets/huggingface/documentation-images/discussions/614), and a link to the well-documented PE-AV collection for checkpoints and end-to-end usage.
I noticed #45612 already proposes a doc fill-in for this page. Happy to drop/adapt this commit if the existing PR is preferred, but the other three fixes are independent of the doc change.

Testing

Run facebook/pe-av-small with the minimal code below.

import torch
from transformers import AutoModel, AutoProcessor

model = AutoModel.from_pretrained("facebook/pe-av-small").eval()
processor = AutoProcessor.from_pretrained("facebook/pe-av-small")
# #4: AutoProcessor.from_pretrained above no longer emits deprecation warnings.
text_inputs = processor(text=["a photo of a cat", "a person speaking"], return_tensors="pt", padding=True)

# #1: previously crashed with TypeError, now returns torch.Size([2, 1024])
with torch.no_grad():
    out = model.get_text_audio_video_embeds(
        input_ids=text_inputs["input_ids"],
        attention_mask=text_inputs.get("attention_mask"),
    )
print(out.shape)

# #2: error now points to get_*_embeds helpers
try:
    model(**text_inputs)
except ValueError as e:
    print(e)

Code Agent Policy

• [x ] I confirm that this is not a pure code agent PR.

Who can review?

First time so don't hate my if I tag wrong people @zucchini-nlp @stevhliu XD

[zucchini-nlp]

niiice!

[zucchini-nlp]

Thanks, lgtm!

[zucchini-nlp]

run-slow: pe_audio_video

[github-actions]

Workflow Run ⚙️

This comment contains run-slow, running the specified jobs:

models: ["models/pe_audio_video"]
quantizations: []

[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.

[github-actions]

CI Results

Workflow Run ⚙️

Commit Info

Context	Commit	Description
RUN	102f0820	workflow commit (merge commit)
PR	9c817809	branch commit (from PR)
main	6c66de3f	base commit (on main)

✅ No failing test specific to this PR 🎉 👏 !

[stevhliu]

nice, the docs will indeed be handled in the other PR! good to merge once that's dropped here :)

[massimilianoviola]

nice, the docs will indeed be handled in the other PR! good to merge once that's dropped here :)

ok, I'll revert the doc change then! thanks

[github-actions]

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

run-slow: pe_audio_video