sgl-project/sglang · error · ValueError

reference audio start time must be non-negative

Error message

reference audio start time must be non-negative

What it means

_load_waveform requires the audio clip start time to be a non-negative finite float. A negative start time would seek before the beginning of the media and is rejected before ffmpeg is invoked.

Source

Thrown at python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/reference_encoding.py:232

) -> tuple[torch.Tensor, int]:
    """Apply the audio material chain.

    Pure-audio references preserve their source rate while normalizing to
    stereo. Video-bearing references first extract 44.1 kHz stereo PCM. The
    audio VAE boundary then performs the single 32 kHz resample below. ffmpeg
    writes bounded interleaved float PCM directly to stdout, avoiding a
    temporary lossless file plus a second decode.
    """

    import numpy as np

    if max_duration_seconds is not None:
        max_duration_seconds = float(max_duration_seconds)
        if not math.isfinite(max_duration_seconds) or max_duration_seconds <= 0:
            raise ValueError("reference audio duration bound must be positive")
    start_time_seconds = float(start_time_seconds)
    if not math.isfinite(start_time_seconds) or start_time_seconds < 0:
        raise ValueError("reference audio start time must be non-negative")

    if material_chain == "audio":
        if source_sample_rate is None or int(source_sample_rate) <= 0:
            raise ValueError("reference audio sample rate must be positive")
        source_rate = int(source_sample_rate)
    elif material_chain in {
        "video.reference_preserve",
        "video_audio.reference_preserve",
    }:
        source_rate = 44100
    else:
        raise ValueError(
            f"unsupported MiniMax H3 audio material chain {material_chain!r}"
        )

    command = [
        "ffmpeg",
        "-v",

View on GitHub (pinned to 0132848349)

Solutions

1. Clamp start times: start = max(0.0, start)
2. Treat negative offsets as offsets from the end and convert: start = max(0.0, total_duration + negative_offset)
3. Validate request timestamps server-side before the pipeline

Example fix

// before
start = marker_t - 5.0  # may go negative

// after
start = max(0.0, marker_t - 5.0)

def valid_start(t) -> bool:
    import math
    return math.isfinite(t) and t >= 0

When it happens

Trigger: Passing start_time_seconds < 0 or NaN to minimax_h3_encode_reference_audio_rows, e.g. from a sign error in timestamp math.

Common situations: Clip offsets computed as (marker - duration) underflowing below zero, or API users sending negative offsets to trim from the end.

Related errors

• MiniMax H3 audio material has no usable sample rate
• MiniMax H3 audio material has no usable channel count
• reference audio duration bound must be positive
• reference audio sample rate must be positive
• reference audio is empty: {audio_path}

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/111dcf3bcca0fa5d. Report an issue: GitHub.