Skip to content

vllm.model_executor.models.moss_transcribe_diarize

Inference-only MOSS-Transcribe-Diarize ASR model.

The checkpoint layout is:

  • model.whisper_encoder.*: Whisper-medium encoder weights.
  • model.vq_adaptor.*: 4x time-merge projector.
  • model.language_model.*: Qwen3-0.6B decoder weights.

Classes:

MossTranscribeDiarizeAudioInputs

Bases: TensorSchema

Dimensions
  • c: Audio chunks
  • m: Mel bins
  • f: Mel frames
  • n: Number of audio items
Source code in vllm/model_executor/models/moss_transcribe_diarize.py 
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
class MossTranscribeDiarizeAudioInputs(TensorSchema):
    """
    Dimensions:
        - c: Audio chunks
        - m: Mel bins
        - f: Mel frames
        - n: Number of audio items
    """

    type: Literal["audio_features"] = "audio_features"

    input_features: Annotated[
        torch.Tensor | None,
        TensorShape("c", "m", "f"),
    ]
    audio_feature_lengths: Annotated[
        torch.Tensor | None,
        TensorShape("c"),
    ]
    audio_chunk_counts: Annotated[
        torch.Tensor | None,
        TensorShape("n"),
    ]

MossTranscribeDiarizeEmbeddingInputs

Bases: TensorSchema

Dimensions
  • n: Number of audio items
  • t: Number of audio tokens
  • h: Hidden size
Source code in vllm/model_executor/models/moss_transcribe_diarize.py 
105
106
107
108
109
110
111
112
113
114
115
116
117
118
class MossTranscribeDiarizeEmbeddingInputs(TensorSchema):
    """
    Dimensions:
        - n: Number of audio items
        - t: Number of audio tokens
        - h: Hidden size
    """

    type: Literal["audio_embeds"] = "audio_embeds"

    audio_embeds: Annotated[
        list[torch.Tensor],
        TensorShape("n", "t", "h", dynamic_dims={"t"}),
    ]