sgl-project/sglang · error · ValueError

Error while loading data {data_str}: {e}

Error message

Error while loading data {data_str}: {e}

What it means

_load_single_item wraps known client-side media loading failures (CLIENT_MEDIA_EXCEPTIONS — HTTP errors, decode failures, missing files) into a ValueError with a truncated (100 char) repr of the offending data, chaining the original exception. It marks the failure as a caller-input problem rather than an internal bug.

Source

Thrown at python/sglang/srt/multimodal/processors/base_processor.py:876

                img, _ = load_image(data, cls.gpu_image_decode)
                if isinstance(img, torch.Tensor):
                    return img  # JPEG already decoded on GPU by nvJPEG
                # PIL decodes lazily; do it here in the io worker so the decode
                # doesn't run later on the event-loop thread.
                if discard_alpha_channel and img.mode != "RGB":
                    return img.convert("RGB")
                img.load()
                return img
            elif modality == Modality.VIDEO:
                return load_video(data, frame_count_limit)
            elif modality == Modality.AUDIO:
                return load_audio(data, audio_sample_rate)

        except CLIENT_MEDIA_EXCEPTIONS as e:
            data_str = str(data)
            if len(data_str) > 100:
                data_str = data_str[:100] + "..."
            raise ValueError(f"Error while loading data {data_str}: {e}") from e
        except Exception as e:
            data_str = str(data)
            if len(data_str) > 100:
                data_str = data_str[:100] + "..."
            raise RuntimeError(f"Error while loading data {data_str}: {e}") from e

    @staticmethod
    def _get_preprocessed_input_format(data):
        """returns the detailed format if the provided data is already preprocessed.
        returns none if the provided data is not preprocessed
        """
        if not isinstance(data, dict):
            return None
        data_format = data.get("format")
        if isinstance(data_format, MultimodalInputFormat):
            return data_format
        if data_format in (
            MultimodalInputFormat.PROCESSOR_OUTPUT.name,

View on GitHub (pinned to 0132848349)

Solutions

1. Check the chained cause (e.__cause__) for the real network/decode error and fix the source
2. Verify the media URL/path/base64 is valid and reachable before submitting the request
3. Retry with backoff only if the cause is transient (e.g. HTTP 503); otherwise correct the data

Example fix

// before
processor.process_mm_data_async(items=[bad_url], ...)
// after
assert requests.head(bad_url, timeout=5).status_code == 200
processor.process_mm_data_async(items=[bad_url], ...)

import requests
for u in urls:
    r = requests.head(u, timeout=5, allow_redirects=True)
    assert r.status_code == 200, f"unreachable media: {u}"

try:
    await processor.process_mm_data_async(...)
except ValueError as e:
    if "Error while loading data" in str(e):
        return client_error(str(e), cause=e.__cause__)  # map to 4xx
    raise

When it happens

Trigger: Loading an image/audio/video item whose URL 404s, whose bytes fail to decode, or whose local path doesn't exist — i.e. any client-classified exception inside the per-item loader.

Common situations: Expired presigned URLs; base64 payloads with header mistakes; unsupported codecs; wrong local paths in batch jobs.

Related errors

• An exception occurred while loading {modality.name} data at
• Unsupported image type: {type(image)}
• When using multiple prompts with multiple input images, prov
• {key}.position_ids is required
• {key}.{field} is required

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/842799566b69fb53. Report an issue: GitHub.