sgl-project/sglang · error · ValueError

audio must be a 2D tensor, but got {self.audio.ndim}D tensor

Error message

audio must be a 2D tensor, but got {self.audio.ndim}D tensor

What it means

When audio is passed directly as a torch.Tensor (not a tuple), it must be 2D — typically shape (channels/batches, time). A 1D or 3D tensor is rejected in __post_init__.

Source

Thrown at python/sglang/srt/multimodal/processors/mimo_audio.py:71

        if isinstance(self.audio, tuple):
            if (
                len(self.audio) != 2
                or not isinstance(self.audio[0], torch.Tensor)
                or not isinstance(self.audio[1], (int, float))
            ):
                raise ValueError(
                    f"audio must be a tuple of (waveform-T, original_sr-int/float), but got {len(self.audio)} elements and {type(self.audio[0])} and {type(self.audio[1])}"
                )
            if self.audio[0].ndim != 1:
                raise ValueError(
                    f"waveform must be a 1D tensor, but got {self.audio[0].ndim}D tensor"
                )
            if self.audio[1] <= 0:
                raise ValueError(
                    f"original_sr must be a positive number, but got {self.audio[1]}"
                )
        if isinstance(self.audio, torch.Tensor) and self.audio.ndim != 2:
            raise ValueError(
                f"audio must be a 2D tensor, but got {self.audio.ndim}D tensor"
            )


class MiMoAudioPipeline:
    """Stateful audio preprocessing pipeline.

    Composable: held by both MiMoProcessor (multimodal) and MiMoV2ASRProcessor.
    Owns the mel spectrogram, resampler cache, http session, and the special
    token ids for ``<|sosp|> <|empty|>* <|eosp|>`` placeholders.
    """

    def __init__(
        self,
        *,
        audio_token_id: int,
        audio_start_token_id: int,
        audio_end_token_id: int,

View on GitHub (pinned to 0132848349)

Solutions

1. For a 1D waveform with a sample rate, use the tuple form: (waveform_1d, original_sr)
2. Otherwise reshape the tensor to 2D with waveform.unsqueeze(0)
3. Check tensor.ndim == 2 before constructing the input

Example fix

# before
audio = waveform_1d  # ndim == 1 -> error
# after
audio = (waveform_1d, 16000)  # tuple form accepts 1D waveform

import torch
if isinstance(audio, torch.Tensor):
    assert audio.ndim == 2, f"tensor audio must be 2D, got {audio.ndim}D"

def is_valid_tensor_audio(a) -> bool:
    import torch
    return not isinstance(a, torch.Tensor) or a.ndim == 2

When it happens

Trigger: Constructing the audio input with a 1D tensor of shape (T,) or a 3D tensor instead of the required 2D layout.

Common situations: User loads mono audio as a 1D tensor and passes it directly; the intended form for raw 1D waveforms is the (waveform, sr) tuple, which is the common mix-up.

Related errors

• waveform must be a 1D tensor, but got {self.audio[0].ndim}D
• audio must be a str, bytes, tuple, torch.Tensor, or np.ndarr
• audio must be a tuple of (waveform-T, original_sr-int/float)
• original_sr must be a positive number, but got {self.audio[1
• torchaudio is required for audio inputs; install torchaudio

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6cc14557d986558e. Report an issue: GitHub.