Fix TextToAudioPipeline crash for tokenizer-only models

[jiqing-feng]

Overview

Any text-to-audio model without a processor is broken on main:

from transformers import pipeline
pipeline(task="text-to-audio", model="facebook/mms-tts-eng")("This is a test")
# TypeError: BatchEncoding.to() got an unexpected keyword argument 'dtype'

tests/pipelines/test_pipelines_text_to_audio.py::TextToAudioPipelineTests::test_vits_model_pt fails on main.

Cause

#40546 (VibeVoice) added an unconditional cast at the end of TextToAudioPipeline.preprocess:

output = output.to(dtype=self.model.dtype)

Models with a processor get a BatchFeature, whose to() accepts dtype. Models without one (VITS, etc.) go through a plain tokenizer and get a BatchEncoding, whose to() only accepts a device.

Fix

Only cast when the preprocessor returned a BatchFeature. A BatchEncoding holds integer tensors only, and BatchFeature.to casts floating point tensors only, so this is a no-op for the tokenizer path either way.

if isinstance(output, BatchFeature):
    output = output.to(dtype=self.model.dtype)

Verification

• pytest tests/pipelines/test_pipelines_text_to_audio.py -k "vits or speecht5" — passes (fails on main).
• facebook/mms-tts-eng, float16, CPU: pipeline runs, returns 45312 samples @ 16 kHz.

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33735433068:2
Result: success | Jobs: 1 | Tests: 1 | Failures: 0 | Duration: 2s