huggingface/transformers · error · ValueError

Error loading audio: {e}

Error message

Error loading audio: {e}

What it means

After massaging the DeepSpeed config against TrainingArguments, `fill_match` recorded every key where the user's DeepSpeed config value conflicts with the Trainer-computed value (batch size, gradient accumulation, optimizer/scheduler params like `total_num_steps`, `warmup_num_steps`, learning rate, etc.). Non-empty `mismatches` abort training with this aggregated ValueError; the recommended remedy baked into the message is to set those DeepSpeed entries to `'auto'` so transformers fills them.

Source

Thrown at src/transformers/audio_utils.py:375

        if force_mono and audio_array.ndim != 1:
            audio_array = audio_array.mean(axis=1)

        buffer = io.BytesIO()
        sf.write(buffer, audio_array, sampling_rate, format=audio_format.upper())
        buffer.seek(0)

        if return_format == "buffer":
            return buffer
        elif return_format == "base64":
            return base64.b64encode(buffer.read()).decode("utf-8")
        elif return_format == "dict":
            return {
                "data": base64.b64encode(buffer.read()).decode("utf-8"),
                "format": audio_format.lower(),
            }

    except Exception as e:
        raise ValueError(f"Error loading audio: {e}")


def conv1d_output_length(module: "torch.nn.Conv1d", input_length: int) -> int:
    """
    Computes the output length of a 1D convolution layer according to torch's documentation:
    https://docs.pytorch.org/docs/stable/generated/torch.nn.Conv1d.html
    """
    return int(
        (input_length + 2 * module.padding[0] - module.dilation[0] * (module.kernel_size[0] - 1) - 1)
        / module.stride[0]
        + 1
    )


def is_valid_audio(audio):
    return (
        is_numpy_array(audio)
        or is_torch_tensor(audio)

View on GitHub (pinned to a597f97485)

Solutions

1. Set the mismatched DeepSpeed keys (listed in the error text) to `"auto"` so they are derived from TrainingArguments
2. Or make the DeepSpeed JSON agree exactly with TrainingArguments values
3. Re-check `num_training_steps` inputs (dataset size, epochs, max_steps) if scheduler params were hardcoded

Example fix

// before (ds_config.json vs TrainingArguments(batch_size=16))
"train_micro_batch_size_per_gpu": 8,
"scheduler": { "params": { "total_num_steps": 100, "warmup_num_steps": 10 } }

// after
"train_micro_batch_size_per_gpu": "auto",
"scheduler": { "params": { "total_num_steps": "auto", "warmup_num_steps": "auto" } }

# mirror the trainer's check before launching
mismatches = []
if ds.get("train_micro_batch_size_per_gpu") not in ("auto", args.per_device_train_batch_size):
    mismatches.append("train_micro_batch_size_per_gpu")
if ds.get("gradient_accumulation_steps") not in ("auto", args.gradient_accumulation_steps):
    mismatches.append("gradient_accumulation_steps")
assert not mismatches, f"fix or set 'auto': {mismatches}"

try:
    trainer.train()
except ValueError as e:
    if "mismatch TrainingArguments values" in str(e):
        # set the listed keys to "auto" in ds_config.json and retry
        raise SystemExit("edit ds_config.json: set listed keys to 'auto'")
    raise

When it happens

Trigger: Passing a DeepSpeed config with hardcoded `train_micro_batch_size_per_gpu`, `gradient_accumulation_steps`, `scheduler.params.total_num_steps`, or `warmup_num_steps` that disagree with `TrainingArguments` (per_device_train_batch_size, gradient_accumulation_steps, computed num_training_steps from dataset length/epochs), or vice versa — changing TrainingArguments after writing the JSON.

Common situations: Reusing one ds_config.json across experiments while changing batch size/epochs in TrainingArguments; tutorials with fixed scheduler values; scripts where dataloader length changed (different dataset or max_steps) invalidating a previously correct total_num_steps.

Related errors

• Invalid return_format: {return_format}. Must be 'base64', 'd
• File not found: {audio}
• This tokenizer does not have a mask token which is necessary
• mlm_probability should be between 0 and 1.
• batched_mm experts dispatch does not support activation_sche

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/6db113e103f88d23. Report an issue: GitHub.