[Model] Refactor Ultravox to use merged input processor

[Isotr0py]

• Refactor Ultravox to use merged input processor
• Ultravox placeholder will be changed to <|audio|> to keep align with HF.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[DarkLight1337]

cc @petersalas regarding the change of placeholder token.

[DarkLight1337]

WDYT of making sampling_rate part of mm_processor_kwargs to make the input format consistent with HF? Even so, we should maintain backwards compatibility for a while.

[Isotr0py]

WDYT of making sampling_rate part of mm_processor_kwargs to make the input format consistent with HF?

But the whisper feature extraction is using a fixed sampling rate, so if we expose the sampling rate to be dynamic, this may cause unnecessary exception.

For example, if we specify the sampling_rate=32000, the ultravox processor will raise an error due to incorrect sampling rate.

from transformers import AutoProcessor
import librosa

processor = AutoProcessor.from_pretrained("fixie-ai/ultravox-v0_3", trust_remote_code=True)
audio, sr = librosa.load("translate_to_chinese.wav")
processor(text="<|audio|>", audio=audio, sampling_rate=32000)

ValueError: The model corresponding to this feature extractor: WhisperFeatureExtractor was trained using a sampling rate of 16000. Please make sure that the provided `raw_speech` input was sampled with 16000 and not 32000.

[DarkLight1337]

WDYT of making sampling_rate part of mm_processor_kwargs to make the input format consistent with HF?

But the whisper feature extraction is using a fixed sampling rate, so if we expose the sampling rate to be dynamic, this may cause unnecessary exception.

For example, if we specify the sampling_rate=32000, the ultravox processor will raise an error due to incorrect sampling rate.

from transformers import AutoProcessor
import librosa

processor = AutoProcessor.from_pretrained("fixie-ai/ultravox-v0_3", trust_remote_code=True)
audio, sr = librosa.load("translate_to_chinese.wav")
processor(text="<|audio|>", audio=audio, sampling_rate=32000)

ValueError: The model corresponding to this feature extractor: WhisperFeatureExtractor was trained using a sampling rate of 16000. Please make sure that the provided `raw_speech` input was sampled with 16000 and not 32000.

I see, so the sampling_rate parameter actually refers to the input, not the HF processor. Let's keep this as is then.

[DarkLight1337]

There seems to be some problem with online inference of this model, please fix it.

[DarkLight1337]

The tests pass so LGTM!