huggingface/transformers · error · ValueError

File not found: {audio}

Error message

File not found: {audio}

What it means

During `HfTrainerDeepSpeedConfig.fill_matches`/`fill_only` auto-filling, transformers resolves `auto` entries that depend on hidden size (e.g. `zero_optimization.reduce_bucket_size`, `stage3_prefetch_bucket_size`) by reading `model.config.hidden_size`, `hidden_sizes`, or the nested `text_config` equivalents. If none of these attributes exist on the model config (non-standard architecture), hidden_size stays None and auto-fill cannot proceed, so it raises with the list of affected keys.

Source

Thrown at src/transformers/audio_utils.py:342

            - `str`: Base64 encoded audio data (if return_format="base64")
            - `dict`: Dictionary with 'data' (base64 encoded audio data) and 'format' keys (if return_format="dict")
            - `io.BytesIO`: BytesIO object containing audio data (if return_format="buffer")
    """
    requires_backends(load_audio_as, ["librosa"])

    if return_format not in ["base64", "dict", "buffer"]:
        raise ValueError(f"Invalid return_format: {return_format}. Must be 'base64', 'dict', or 'buffer'")

    try:
        # Load audio bytes from URL or file
        audio_bytes = None
        if audio.startswith(("http://", "https://")):
            audio_bytes = _fetch_audio_bytes(audio, timeout=timeout)
        elif os.path.isfile(audio):
            with open(audio, "rb") as audio_file:
                audio_bytes = audio_file.read()
        else:
            raise ValueError(f"File not found: {audio}")

        # Process audio data
        with io.BytesIO(audio_bytes) as audio_file:
            with sf.SoundFile(audio_file) as f:
                audio_array = f.read(dtype="float32")
                original_sr = f.samplerate
                audio_format = f.format
                if sampling_rate is not None and sampling_rate != original_sr:
                    # Resample audio to target sampling rate
                    audio_array = soxr.resample(audio_array, original_sr, sampling_rate, quality="HQ")
                else:
                    sampling_rate = original_sr

        # Convert to mono if needed
        if force_mono and audio_array.ndim != 1:
            audio_array = audio_array.mean(axis=1)

        buffer = io.BytesIO()

View on GitHub (pinned to a597f97485)

Solutions

1. Replace the `auto` values for the listed keys with explicit integers in your DeepSpeed config (e.g. `reduce_bucket_size: 5000000`, `stage3_prefetch_bucket_size: 4500000`)
2. Or expose `hidden_size`/`hidden_sizes` on your custom config class so auto-fill works
3. Or route the config through a `text_config` attribute carrying `hidden_size` if the model wraps a text backbone

Example fix

// before (ds_config.json)
"zero_optimization": { "stage": 3, "reduce_bucket_size": "auto", "stage3_prefetch_bucket_size": "auto" }

// after
"zero_optimization": { "stage": 3, "reduce_bucket_size": 5000000, "stage3_prefetch_bucket_size": 4500000 }

cfg = model.config
hidden = getattr(cfg, "hidden_size", None) or (
    max(getattr(cfg, "hidden_sizes", [])) if hasattr(cfg, "hidden_sizes") else None
) or getattr(getattr(cfg, "text_config", None), "hidden_size", None)
if hidden is None:
    # replace 'auto' bucket keys with explicit ints before training
    for k in ("zero_optimization.reduce_bucket_size", "zero_optimization.stage3_prefetch_bucket_size"):
        if deepspeed_dict_get(ds_cfg, k) == "auto":
            raise SystemExit(f"set {k} to an integer; model config has no hidden_size")

When it happens

Trigger: Training with DeepSpeed ZeRO (config containing `auto` bucket-size keys) a model whose config class lacks `hidden_size`/`hidden_sizes` (and `text_config.hidden_size[s]`) — e.g. some vision/multimodal/audio models or heavily custom configs.

Common situations: Custom model architectures; older multimodal configs without text_config; wrapper configs that hide the LM config under a different attribute name.

Related errors

• Error loading audio: {e}
• Invalid return_format: {return_format}. Must be 'base64', 'd
• out_indices must be a list, got {type(self._out_indices)}
• out_indices must be valid indices for stage_names {self.stag
• out_indices must not contain any duplicates, got {self._out_

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/e4873e7d55f8bacf. Report an issue: GitHub.