sgl-project/sglang · error · ValueError

Incorrect type of image crop. Got type: {type(images_crop)}

Error message

Incorrect type of image crop. Got type: {type(images_crop)}

What it means

Raised by _parse_and_validate_image_input in deepseek_ocr.py when images_crop (the crop-region tensor) is neither torch.Tensor nor list. The three-part validation (pixel_values, images_crop, images_spatial_crop) must all pass before the encoder runs.

Source

Thrown at python/sglang/srt/models/deepseek_ocr.py:1615

            if not has_images:
                return None
        elif torch.sum(pixel_values).item() == 0:
            return None

        if pixel_values is not None:
            if not isinstance(pixel_values, (torch.Tensor, list)):
                raise ValueError(
                    "Incorrect type of pixel values. " f"Got type: {type(pixel_values)}"
                )

            if not isinstance(images_spatial_crop, (torch.Tensor, list)):
                raise ValueError(
                    "Incorrect type of image sizes. "
                    f"Got type: {type(images_spatial_crop)}"
                )

            if not isinstance(images_crop, (torch.Tensor, list)):
                raise ValueError(
                    "Incorrect type of image crop. " f"Got type: {type(images_crop)}"
                )

            return [pixel_values, images_crop, images_spatial_crop]

        raise AssertionError("This line should be unreachable.")

    def _pixel_values_to_embedding(
        self,
        pixel_values: torch.Tensor,
        images_crop: torch.Tensor,
        images_spatial_crop: torch.Tensor,
        has_local_crops: Optional[List[bool]] = None,
    ) -> NestedTensors:

        # Pixel_values (global view): [n_image, batch_size, 3, height, width]
        # images_spatial_crop: [n_image, batch_size, [num_tiles_w, num_tiles_h]]
        # images_crop (local view): [n_image, batch_size, num_pathes, 3, h, w]

View on GitHub (pinned to 0132848349)

Solutions

1. Pass processor-generated images_crop as a tensor/list
2. Ensure the whole multimodal kwargs dict comes from one processor(...) call rather than being assembled piecemeal
3. Upgrade/downgrade the processor to the version paired with the sglang model implementation

Example fix

// before
inputs = {"pixel_values": pv, "images_spatial_crop": sc, "images_crop": crop_dict}

// after
inputs = {"pixel_values": pv, "images_spatial_crop": sc,
          "images_crop": torch.tensor(crop_dict["boxes"])}

crop = mm_kwargs.get("images_crop")
assert isinstance(crop, (torch.Tensor, list)), "images_crop must be tensor/list"

def is_valid_images_crop(v) -> bool:
    return isinstance(v, (torch.Tensor, list))

When it happens

Trigger: get_multimodal_embeddings is called with images_crop set to an unexpected type — numpy array, dict from an unserialized processor output, string, or a scalar — failing the isinstance((torch.Tensor, list)) check at deepseek_ocr.py:1615.

Common situations: Processor/model version drift where the crop tensor key is absent and defaults to a placeholder, JSON-serialized multimodal payloads that decode crops as dicts/lists-of-dicts, or custom OCR clients that only populate pixel data.

Related errors

• Incorrect type of pixel values. Got type: {type(pixel_values
• Incorrect type of image sizes. Got type: {type(images_spatia
• {field_name} must be a tensor, list of tensors, list of sequ
• Image aspect ratio must be smaller than 200
• You have to specify pixel_values or pixel_embeds

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/9984ba25f3aac570. Report an issue: GitHub.