huggingface/transformers · error · ValueError

not supported filetype

Error message

not supported filetype

What it means

`deepgemm_fp8_fp4_experts_forward` refuses to run when the model spans more than one CUDA device in a single process (`_deepgemm_disabled` was set at load time by the FP8 quantizer). DeepGEMM kernels are bound to a single CUDA context and would corrupt results across devices, so multi-GPU `device_map='auto'` + DeepGEMM experts is hard-disabled at load.

Source

Thrown at src/transformers/audio_utils.py:173

        b"caff": "caf",
        b".snd": "au",
        b"#!AMR": "amr",
        b"wvpk": "wv",
        b".SoX": "sox",
        b"XoS.": "sox",
        b"Creative Voice File": "voc",
        b"\x64\xa3\x01\x00": "sf",
        b"\x00\x01\xa3\x64": "sf",
        b"\x0b\x77": "ac3",
        b"\x00\x00\x01\xba": "mpg",
        b"FLV": "flv",
        b"ID3": "mp3",
    }
    for sig, filetype in signatures.items():
        if head.startswith(sig):
            return filetype

    raise ValueError("not supported filetype")


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:

View on GitHub (pinned to a597f97485)

Solutions

1. Use true multi-process parallelism (TP/EP via accelerate: `accelerate launch` with a parallelism config, one device per process)
2. Or run the model on a single GPU (enough memory) so DeepGEMM stays enabled
3. Or explicitly choose `experts_implementation='grouped_mm'` which is safe across devices in one process

Example fix

# before
model = AutoModelForCausalLM.from_pretrained(
    "deepseek-ai/DeepSeek-V3", device_map="auto"  # 2+ GPUs in one process
)  # -> RuntimeError at forward

# after
# option A: accelerate TP/EP (one GPU per process)
# accelerate launch --num_processes 8 run_infer.py
# option B: grouped_mm experts
model.set_experts_implementation("grouped_mm")

cuda_ids = torch.cuda.device_count()
if cuda_ids > 1:  # single process, multiple devices
    experts_impl = "grouped_mm"            # safe in-process
    # or plan real TP/EP with accelerate instead of device_map="auto"
device_map = "auto" if cuda_ids == 1 else None

try:
    out = model(input_ids)
except RuntimeError as e:
    if "multiple CUDA devices" in str(e):
        model.set_experts_implementation("grouped_mm")
        out = model(input_ids)
    else:
        raise

When it happens

Trigger: `from_pretrained(..., device_map='auto')` sharding a DeepSeek-style FP4/FP8 MoE model over 2+ GPUs in one process, with the deepgemm experts implementation selected (default for supported checkpoints).

Common situations: Fitting a 671B model on 2–8 GPUs via device_map auto on a single node instead of tensor/expert parallel; notebooks pipelines that shard across visible GPUs.

Related errors

• function {activation_string} not found in ACT2FN mapping {li
• Incorrect audio source. Must be a valid URL starting with `h
• Incorrect format used for `audio`. Should be a numpy array o
• Unknown backend {backend!r}; expected 'auto', 'torchcodec',
• The audio source is a '{filetype}' file, which librosa canno

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/0e4d1ddcfa5d7c45. Report an issue: GitHub.