docling-project/docling · error · TypeError

Expected scalar-like tensor with one element, got shape={tup

Error message

Expected scalar-like tensor with one element, got shape={tuple(value.shape)}

What it means

Raised as TypeError by HfVisionModelMixin._as_float when a torch.Tensor score does not contain exactly one element (numel() != 1). Same contract as the ndarray variant: score values must be scalar-like tensors before conversion to float.

Source

Thrown at docling/models/inference_engines/common/hf_vision_base.py:128

        return self._id_to_label

    @staticmethod
    def _as_float(value: Any) -> float:
        if isinstance(value, Real):
            return float(value)

        if isinstance(value, np.ndarray):
            if value.size != 1:
                raise TypeError(
                    f"Expected scalar-like ndarray with size 1, got shape={value.shape}"
                )
            return float(value.reshape(-1)[0])

        import torch

        if isinstance(value, torch.Tensor):
            if value.numel() != 1:
                raise TypeError(
                    f"Expected scalar-like tensor with one element, got shape={tuple(value.shape)}"
                )
            return float(value.item())

        raise TypeError(f"Unsupported score value type: {type(value)!r}")

    @staticmethod
    def _as_int(value: Any) -> int:
        if isinstance(value, Integral):
            return int(value)

        if isinstance(value, np.ndarray):
            if value.size != 1:
                raise TypeError(
                    f"Expected scalar-like ndarray with size 1, got shape={value.shape}"
                )
            return int(value.reshape(-1)[0])

View on GitHub (pinned to 61d76f1ff3)

Solutions

1. Reduce the tensor to one element first: scores[i, labels[i]] or scores.max().
2. Call .item() yourself when you know it is scalar, or squeeze and assert numel()==1 in your post-processor.
3. Align your custom head output with the expected [N] scalar-score layout.

Example fix

# before
conf = model._as_float(scores_tensor[i])  # shape (num_classes,) -> TypeError

# after
conf = model._as_float(scores_tensor[i, labels[i]])  # numel()==1

if torch.is_tensor(score):
    assert score.numel() == 1, f'score tensor must have 1 element, got {tuple(score.shape)}'

import torch

def is_scalar_tensor(value) -> bool:
    return not torch.is_tensor(value) or value.numel() == 1

When it happens

Trigger: Passing a multi-element torch tensor (e.g. a [num_classes] score row or [1, num_boxes] slice) to _as_float.

Common situations: Feeding raw model logits rows into result construction without argmax/max reduction; batch post-processing that forgot to index per detection.

Related errors

• Expected scalar-like ndarray with size 1, got shape={value.s
• Unsupported score value type: {type(value)!r}
• Unsupported label value type: {type(value)!r}
• Unsupported input type: {type(self.path_or_stream)}
• Unexpected: {type(self.path_or_stream)=}

AI-assisted analysis of docling-project/docling@61d76f1ff3 (2026-08-14). Data as JSON: /api/errors/e665d27113ce054b. Report an issue: GitHub.