vllm.transformers_utils.processors.voxtral ¶

class MistralCommonFeatureExtractor:
    """
    Provide a HF-compatible interface for
    `mistral_common.tokens.tokenizers.multimodal.AudioEncoder`.
    """

    def __init__(self, audio_encoder: AudioEncoder) -> None:
        self.audio_encoder = audio_encoder

    @property
    def sampling_rate(self):
        return self.audio_encoder.audio_config.sampling_rate

    @property
    def frame_rate(self):
        return self.audio_encoder.audio_config.frame_rate

    def __call__(
        self,
        audios: AudioInput,
        return_tensors: str | TensorType | None = None,
        **kwargs,
    ) -> BatchFeature:
        audios_lst = [audios] if not isinstance(audios, list) else audios

        audios_processed = list[torch.Tensor]()

        for audio in audios_lst:
            audio = np.asarray(audio, dtype=np.float32).ravel()
            if not self.audio_encoder.audio_config.is_streaming:
                audio = self.audio_encoder.pad(audio, self.sampling_rate)

            audios_processed.append(torch.tensor(audio))

        return BatchFeature(
            {"audio_arrays": audios_processed}, tensor_type=return_tensors
        )

    def get_num_audio_tokens(self, audio_length: int) -> int:
        return ceil(audio_length / (self.sampling_rate // self.frame_rate))