huggingface/transformers · error · ValueError

Incorrect audio source. Must be a valid URL starting with `h

Error message

Incorrect audio source. Must be a valid URL starting with `http://` or `https://`, a valid path to an audio file, or a base64 encoded string. Got {audio}. Failed with {e}

What it means

The DeepGEMM FP8/FP4 experts dispatch only supports dynamic (per-token) activation quantization; when the experts module was quantized with `activation_scheme='static'` (calibrated per-tensor input scales), the dispatch raises `NotImplementedError` on the first forward. DeepGEMM's kernels need per-row scale factors, which static schemes do not provide.

Source

Thrown at src/transformers/audio_utils.py:192


def _resolve_audio_source(audio: str, timeout: float | None = None) -> "str | bytes":
    """Resolve an audio source string to a local file path or raw bytes for a decoder.

    Accepts `http(s)://` URLs (fetched with retry), local file paths (returned unchanged),
    and base64 strings (optionally wrapped as a `data:...` URI).
    """
    if audio.startswith(("http://", "https://")):
        return _fetch_audio_bytes(audio, timeout=timeout)
    if os.path.isfile(audio):
        return audio
    # Not a URL or a local path — assume base64, optionally wrapped as a `data:<media-type>;base64,` URI
    if audio.startswith("data:"):
        audio = audio.split(",", 1)[1]
    try:
        return base64.b64decode(audio)
    except Exception as e:
        raise ValueError(
            "Incorrect audio source. Must be a valid URL starting with `http://` or `https://`, "
            f"a valid path to an audio file, or a base64 encoded string. Got {audio}. Failed with {e}"
        )


def load_audio(audio: str | np.ndarray, sampling_rate=16000, timeout=None, backend: str = "auto") -> np.ndarray:
    """
    Loads `audio` to an np.ndarray object.

    Args:
        audio (`str` or `np.ndarray`):
            The audio to be loaded to the numpy array format. If a `str`, it can be an `http(s)://`
            URL, a local file path, or a base64-encoded string (optionally wrapped as a
            `data:<media-type>;base64,` URI).
        sampling_rate (`int`, *optional*, defaults to 16000):
            The sampling rate to be used when loading the audio. It should be same as the
            sampling rate the model you will be using further was trained with.
        timeout (`float`, *optional*):

View on GitHub (pinned to a597f97485)

Solutions

Switch experts dispatch to `grouped_mm` (or the default) which supports static activation scales
Re-quantize/calibrate the checkpoint with `activation_scheme='dynamic'`
Catch NotImplementedError and fall back per-layer if building a generic runner

Example fix

# before
model.set_experts_implementation("deepgemm")
out = model(x)  # activation_scheme == "static" -> NotImplementedError

# after
model.set_experts_implementation("grouped_mm")

Defensive patterns

Strategy: validation

Validate before calling

scheme = getattr(experts_module, "activation_scheme", None)
if scheme == "static":
    model.set_experts_implementation("grouped_mm")  # deepgemm needs dynamic per-token quant

Try / catch

try:
    out = experts(hidden, idx, w)
except NotImplementedError as e:
    if "static activation quantization" in str(e):
        model.set_experts_implementation("grouped_mm")
        out = model(input_ids)
    else:
        raise

Prevention

Check activation_scheme in the checkpoint's quantization config at load
Prefer dynamic activation FP8 checkpoints for DeepGEMM deployments

When it happens

Trigger: Loading an FP8 MoE checkpoint calibrated with static activation scales (e.g. DeepSeek-V2 static variants, `QuantizerConfig(activation_scheme='static')`) and running `experts_implementation='deepgemm'`.

Common situations: Switching dispatch from the default to 'deepgemm' on an older static-FP8 checkpoint; teams re-using calibrated V2 scales with V3-style kernels.

Related errors

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/d3e525f015878f8c. Report an issue: GitHub.