Skip to content

vllm.model_executor.models.midashenglm

Inference-only MiDashengLM model compatible with HuggingFace weights.

Classes:

Functions:

MiDashengLMAudioInputs

Bases: TensorSchema

Dimensions
  • bn: Batch size * number of audios
  • p: Number of sampling points
Source code in vllm/model_executor/models/midashenglm.py 
513
514
515
516
517
518
519
520
521
522
class MiDashengLMAudioInputs(TensorSchema):
    """

    Dimensions:
        - bn: Batch size * number of audios
        - p: Number of sampling points
    """

    input_values: Annotated[torch.Tensor, TensorShape("n", "p")]
    audio_length: Annotated[torch.Tensor, TensorShape("n")]

calculate_mel_frames_dasheng(audio_length_samples, n_fft=512, hop_size=160, dasheng_subsampling=4, center=True, model_subsampling=5)

Calculate the number of Mel-spectrogram frames.

Source code in vllm/model_executor/models/midashenglm.py 
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
def calculate_mel_frames_dasheng(
    audio_length_samples: int,
    n_fft: int = 512,
    hop_size: int = 160,
    dasheng_subsampling: int = 4,
    center=True,
    model_subsampling: int = 5,
) -> int:
    """Calculate the number of Mel-spectrogram frames."""
    if center:
        audio_length_samples = audio_length_samples + n_fft

    return (
        int(1 + ((audio_length_samples - n_fft) / hop_size))
        // dasheng_subsampling
        // model_subsampling
    )