sgl-project/sglang · error · ValueError

waveform must be a 1D tensor, but got {self.audio[0].ndim}D

Error message

waveform must be a 1D tensor, but got {self.audio[0].ndim}D tensor

What it means

Inside the (waveform, sr) tuple form, the waveform must be a 1D tensor of shape (T,). A 2D/3D tensor (e.g. batched or channel-first audio) is rejected.

Source

Thrown at python/sglang/srt/multimodal/processors/mimo_audio.py:63

    audio: str | bytes | tuple | torch.Tensor | np.ndarray

    def __post_init__(self):
        if not isinstance(self.audio, (str, bytes, tuple, torch.Tensor, np.ndarray)):
            raise ValueError(
                f"audio must be a str, bytes, tuple, torch.Tensor, or np.ndarray, but got {type(self.audio)}"
            )
        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.

View on GitHub (pinned to 0132848349)

Solutions

Reshape to 1D: waveform.squeeze(0) for mono, or mean over the channel dim for stereo
Use waveform.reshape(-1) when the audio is mono stored with a leading 1
Check waveform.ndim == 1 before constructing the input

Example fix

# before
waveform, sr = torchaudio.load("a.wav")  # (channels, T)
audio = (waveform, sr)
# after
waveform, sr = torchaudio.load("a.wav")
audio = (waveform.mean(dim=0), sr)  # mono 1D

Defensive patterns

Strategy: validation

Validate before calling

assert isinstance(audio[0], torch.Tensor) and audio[0].ndim == 1, f"got {audio[0].ndim}D"

Type guard

def is_1d_waveform(w) -> bool:
    import torch
    return isinstance(w, torch.Tensor) and w.ndim == 1

Prevention

Squeeze mono audio after torchaudio.load before passing downstream
Add a shape assertion in your audio loading helper

When it happens

Trigger: Passing a tensor of shape (1, T) or (B, C, T) as the waveform element of the audio tuple.

Common situations: Loading audio via torchaudio.load which returns shape (channels, T); forgetting to squeeze mono audio to 1D.

Related errors

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/8e7d334fd27c96de. Report an issue: GitHub.