huggingface/transformers · error · ValueError

Some items in the output dictionary have a different batch s

Error message

Some items in the output dictionary have a different batch size than others.

What it means

Before per-example truncation/padding, the feature extractor verifies every value in the processed-features dict has the same batch dimension (len == batch size of the main input). Any key whose first dimension differs (a scalar, a single unbatched example, or a ragged list) makes the batch inconsistent and raises this ValueError.

Source

Thrown at src/transformers/feature_extraction_sequence_utils.py:185

                )

        for key, value in processed_features.items():
            if isinstance(value[0], (int, float)):
                processed_features[key] = to_numpy(value)
            elif not isinstance(value, np.ndarray):
                # An already-batched numpy array can be used as-is; splitting it
                # into a list of per-example arrays is pure overhead and is very
                # slow for large inputs (e.g. long audio).
                processed_features[key] = [to_numpy(v) for v in value]

        # Convert padding_strategy in PaddingStrategy
        padding_strategy = self._get_padding_strategies(padding=padding, max_length=max_length)

        required_input = processed_features[self.model_input_names[0]]

        batch_size = len(required_input)
        if not all(len(v) == batch_size for v in processed_features.values()):
            raise ValueError("Some items in the output dictionary have a different batch size than others.")

        truncated_inputs = []
        for i in range(batch_size):
            inputs = {k: v[i] for k, v in processed_features.items()}
            # truncation
            inputs_slice = self._truncate(
                inputs,
                max_length=max_length,
                pad_to_multiple_of=pad_to_multiple_of,
                truncation=truncation,
            )
            truncated_inputs.append(inputs_slice)

        if padding_strategy == PaddingStrategy.LONGEST:
            # make sure that `max_length` cannot be longer than the longest truncated length
            max_length = max(len(input_slice[self.model_input_names[0]]) for input_slice in truncated_inputs)
            padding_strategy = PaddingStrategy.MAX_LENGTH

View on GitHub (pinned to a597f97485)

Solutions

1. Make every key in the dict have exactly batch_size entries along dim 0
2. Remove keys that are not per-example (pass them outside the feature dict)
3. Use the feature extractor's own __call__ on raw audio so all keys are built consistently

Example fix

# before
batch = {"input_values": [a, b], "labels": [lab]}  # len mismatch

# after
batch = {"input_values": [a, b], "labels": [lab0, lab1]}

def assert_consistent_batch(batch: dict):
    sizes = {k: len(v) for k, v in batch.items()}
    if len(set(sizes.values())) > 1:
        raise ValueError(f"inconsistent batch sizes: {sizes}")
    return batch

def is_uniform_batch(batch: dict) -> bool:
    return len({len(v) for v in batch.values()}) == 1

try:
    fe.pad(batch, padding=True)
except ValueError as e:
    if "different batch size" in str(e):
        n = len(batch[fe.model_input_names[0]])
        batch = {k: v[:n] if len(v) > n else v for k, v in batch.items()}
        fe.pad(batch, padding=True)
    else:
        raise

When it happens

Trigger: Mixing batched and unbatched entries: e.g. {'input_values': [a, b, c], 'attention_mask': one_array_of_len_1} or attaching a per-batch scalar/metadata array with a different length; also lists of raw audio where one key has fewer/more examples.

Common situations: Hand-assembled batches where extra keys (labels, custom metadata) don't match the batch size; off-by-one when slicing batches; mixing single-example and multi-example dicts.

Related errors

• You should supply an instance of `transformers.BatchFeature`
• Invalid padding strategy:{padding_side}
• When setting ``padding={PaddingStrategy.MAX_LENGTH}``, make
• Asking to pad but the feature_extractor does not have a padd
• type of {first_element} unknown: {type(first_element)}. Shou

AI-assisted analysis of huggingface/transformers@a597f97485 (2026-08-14). Data as JSON: /api/errors/27b2dfe67a6fcbc2. Report an issue: GitHub.