sgl-project/sglang · error · ValueError

audio_sampling_rate must be set in processor_config or…

Error message

audio_sampling_rate must be set in processor_config or audio_config

What it means

Raised in MiMoProcessor.__init__ when no audio sampling rate can be resolved: processor_config['audio_sampling_rate'] is absent and audio_config provides neither 'sampling_rate' nor 'sample_rate'. The audio feature extractor cannot resample/interpret waveforms without it.

Solutions

1. Set processor_config.audio_sampling_rate (commonly 16000) in preprocessor_config.json or the override passed to MiMoProcessor
2. If an audio_config exists, add sampling_rate under the expected key
3. Use a checkpoint revision that ships complete audio config

Example fix

# before: preprocessor_config.json has no audio key
# after
{"audio_sampling_rate": 16000, "fps": 2.0, ...}  # in preprocessor_config.json

rate = processor_config.get('audio_sampling_rate') or (audio_config or {}).get('sampling_rate') or (audio_config or {}).get('sample_rate')
assert rate is not None, 'set processor_config.audio_sampling_rate (e.g. 16000)'

try:
    proc = MiMoProcessor(hf_config, server_args, _processor)
except ValueError as e:
    if 'audio_sampling_rate' in str(e):
        processor_config['audio_sampling_rate'] = 16000
        proc = MiMoProcessor(hf_config, server_args, _processor)
    else:
        raise

When it happens

Trigger: Loading a MiMo-V2 checkpoint whose preprocessor/processor config lacks audio_sampling_rate while the HF audio_config also has no rate fields, then constructing the processor (i.e., at server/model startup with audio enabled).

Common situations: Checkpoints shipped without audio config sections; configs using a different key name (e.g. 'samplingrate' or nested); older checkpoint revisions predating the audio pipeline.

Understand the failure class

Background: "missing required config value" errors: why libraries refuse to start when a configuration key is empty, unset, or blank — this error's family across 48 libraries.

Related errors

• JoyEcho audio scheduler was not prepared.
• must be a dict-like config, got
• processor_config. must be set for MiMo-V2
• A CUDA VMM-enabled model must provide a multimodal processor
• Activation function is not supported.

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/2fa4fd24f64ff135. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/srt/multimodal/processors/mimo_v2.py:1556

            if (
                rope_scaling.get("type", None) == "default"
                and rope_scaling.get("mrope_section", None) is not None
            ):
                rope_type = "mrope"

        processor_config = self._normalize_config_dict(
            getattr(hf_config, "processor_config", {}), "processor_config"
        )
        audio_config = self._normalize_config_dict(
            getattr(hf_config, "audio_config", None), "audio_config"
        )
        self.audio_sample_rate = processor_config.get("audio_sampling_rate")
        if self.audio_sample_rate is None:
            self.audio_sample_rate = audio_config.get(
                "sampling_rate"
            ) or audio_config.get("sample_rate")
        if self.audio_sample_rate is None:
            raise ValueError(
                "audio_sampling_rate must be set in processor_config or audio_config"
            )

        self.IM_START_TOKEN_ID = self._require_config_value(
            processor_config, "vision_start_token_id"
        )
        self.IM_END_TOKEN_ID = self._require_config_value(
            processor_config, "vision_end_token_id"
        )
        self.IM_TOKEN_ID = self._require_config_value(
            processor_config, "image_token_id"
        )
        self.VIDEO_TOKEN_ID = self._require_config_value(
            processor_config, "video_token_id"
        )
        self.vision_start_token_id = self.IM_START_TOKEN_ID
        self.vision_end_token_id = self.IM_END_TOKEN_ID

View on GitHub (pinned to 0132848349)