Fix Qwen Omni batched text post-processing

[Sunt-ing]

What does this PR do?

Qwen2_5OmniProcessor and Qwen3OmniMoeProcessor decoded generated_outputs[0] for text post-processing. That is correct when generate() returns a tuple-like multimodal result, where the text sequences are the first element. It is wrong for text-only generation, where generate() returns the full (batch, seq) tensor directly.

For batched pipeline("any-to-any") text calls, the model generates a two-row tensor but the processor decodes only the first row, so the pipeline returns one record for a two-item batch.

This keeps the tuple/list path for multimodal outputs and decodes the whole tensor for text-only outputs. The regression guard covers both Omni processors.

End-to-end pipeline repro (real AnyToAnyPipeline + real Qwen3-Omni processor)

Environment:

Hardware: AutoDL c4090, NVIDIA GeForce RTX 4090 present; repro run CPU-only with CUDA_VISIBLE_DEVICES=""
OS: Linux AutoDL container
Python: 3.12
PyTorch: 2.8.0+cu128
Transformers main: b70d02fc724d
This PR head: 651c109d67
Processor metadata: Qwen/Qwen3-Omni-30B-A3B-Instruct via AutoProcessor.from_pretrained
Model: small fake Qwen3OmniMoeForConditionalGeneration; it only replaces 30B weight download and still exercises pipeline("any-to-any") plus the real processor post-processing path
Command: CUDA_VISIBLE_DEVICES="" python repro.py

import torch

from transformers import AutoProcessor, GenerationConfig, pipeline


class Qwen3OmniMoeForConditionalGeneration(torch.nn.Module):
    input_modalities = ("image", "video", "audio", "text")
    output_modalities = ("text", "audio")

    def __init__(self):
        super().__init__()
        self.config = type("Cfg", (), {"_commit_hash": None, "model_type": "qwen3_omni_moe"})()
        self.generation_config = GenerationConfig(max_new_tokens=1)

    @property
    def device(self):
        return torch.device("cpu")

    def can_generate(self):
        return True

    def generate(self, input_ids=None, **kwargs):
        print("GEN_BATCH", tuple(input_ids.shape))
        extra = torch.arange(42, 42 + input_ids.shape[0], dtype=input_ids.dtype).unsqueeze(1)
        return torch.cat([input_ids, extra], dim=1)


processor = AutoProcessor.from_pretrained("Qwen/Qwen3-Omni-30B-A3B-Instruct")
pipe = pipeline("any-to-any", model=Qwen3OmniMoeForConditionalGeneration(), processor=processor)
out = pipe(text=["hello", "world"], return_full_text=False, max_new_tokens=1)
print("PIPE_OUT_LEN", len(out), out)

# main b70d02fc724d
GEN_BATCH (2, 1)
PIPE_OUT_LEN 1 [{'input_text': 'hello', 'generated_text': 'K'}]

# after this PR
GEN_BATCH (2, 1)
PIPE_OUT_LEN 2 [{'input_text': 'hello', 'generated_text': 'K'}, {'input_text': 'world', 'generated_text': 'L'}]

Regression tests and repository checks also passed.

Regression tests and repository checks

Tests:

python -m pytest -q \
  tests/models/qwen2_5_omni/test_processing_qwen2_5_omni.py::test_qwen2_5_omni_post_process_multimodal_output_keeps_text_batch \
  tests/models/qwen3_omni_moe/test_processing_qwen3_omni_moe.py::test_qwen3_omni_moe_post_process_multimodal_output_keeps_text_batch

2 passed in 0.16s

Also ran:

python -m ruff check src/transformers/models/qwen2_5_omni/processing_qwen2_5_omni.py src/transformers/models/qwen3_omni_moe/processing_qwen3_omni_moe.py tests/models/qwen2_5_omni/test_processing_qwen2_5_omni.py tests/models/qwen3_omni_moe/test_processing_qwen3_omni_moe.py
python -m ruff format --check src/transformers/models/qwen2_5_omni/processing_qwen2_5_omni.py src/transformers/models/qwen3_omni_moe/processing_qwen3_omni_moe.py tests/models/qwen2_5_omni/test_processing_qwen2_5_omni.py tests/models/qwen3_omni_moe/test_processing_qwen3_omni_moe.py
python utils/check_modular_conversion.py --files src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py
python utils/check_copies.py --file src/transformers/models/qwen2_5_omni/processing_qwen2_5_omni.py
python utils/check_copies.py --file src/transformers/models/qwen3_omni_moe/processing_qwen3_omni_moe.py
git diff --check

Code Agent Policy

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

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline and the
  Pull Request checks?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes according to the guidelines?
• Did you write any new necessary tests?

Who can review?

cc @zucchini-nlp

[github-actions]

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

run-slow: qwen2_5_omni, qwen3_omni_moe

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 28701194782:2
Result: success | Jobs: 3 | Tests: 140 | Failures: 0 | Duration: 15m 14s