hiyouga/LlamaFactory · error · ValueError

MOSS-VL batch metadata must have one entry per sample.

Error message

MOSS-VL batch metadata must have one entry per sample.

What it means

MOSS-VL batching sanity check in _get_mm_inputs: imglens, vidlens and batch_ids must all have the same length (one entry per sample in the batch). These lists are produced internally by the data pipeline; a mismatch indicates corrupted batch metadata rather than user data.

Source

Thrown at src/llamafactory/data/mm_plugin.py:680

    @override
    def get_mm_inputs(
        self,
        images: list["ImageInput"],
        videos: list["VideoInput"],
        audios: list["AudioInput"],
        imglens: list[int],
        vidlens: list[int],
        audlens: list[int],
        batch_ids: list[list[int]],
        processor: Optional["MMProcessor"],
    ) -> dict[str, Union[list[int], "torch.Tensor"]]:
        self._validate_input(processor, images, videos, audios)
        if audios:
            raise ValueError("MOSS-VL does not support audio inputs.")

        if not (len(imglens) == len(vidlens) == len(batch_ids)):
            raise ValueError("MOSS-VL batch metadata must have one entry per sample.")
        final_pixel_values = []
        final_grid_thw = []
        media_nums_per_sample = []
        image_offset = 0
        video_offset = 0
        for imglen, vidlen, input_ids in zip(imglens, vidlens, batch_ids):
            sample_images = images[image_offset : image_offset + imglen]
            sample_videos = videos[video_offset : video_offset + vidlen]
            image_offset += imglen
            video_offset += vidlen
            image_chunks, image_grids = [], []
            if sample_images:
                regularized_images = self._regularize_images(
                    sample_images,
                    image_max_pixels=2**63 - 1,
                    image_min_pixels=1,
                )["images"]
                image_kwargs = {"return_tensors": "pt"}

View on GitHub (pinned to f28afaf635)

Solutions

1. If calling the API directly, assert len(imglens) == len(vidlens) == len(batch_ids) before the call.
2. Remove local patches to the data collator or update them to the current LlamaFactory version.
3. Reproduce with stock LlamaFactory data pipeline to confirm the bug is in custom code.

Example fix

# before
mm_inputs = plugin._get_mm_inputs(images, videos, audios, imglens, vidlens, audlens, batch_ids, processor)
# after
assert len(imglens) == len(vidlens) == len(batch_ids), 'per-sample metadata must align'
mm_inputs = plugin._get_mm_inputs(images, videos, audios, imglens, vidlens, audlens, batch_ids, processor)

assert len(imglens) == len(vidlens) == len(batch_ids), 'per-sample batch metadata must align'

When it happens

Trigger: Custom code calling plugin._get_mm_images/videos path with hand-built imglens/vidlens/batch_ids of differing lengths; or a patched data collator that returns mismatched lengths.

Common situations: Users subclassing the trainer/collator for multimodal batching and building the lens lists independently; version skew after upgrading LlamaFactory where collator output changed.

Related errors

• MOSS-VL media lengths do not consume all provided inputs.
• MOSS-VL encountered nested video token blocks after tokeniza
• MOSS-VL encountered a video end token without a matching sta
• MOSS-VL encountered an incomplete video token block after to
• MOSS-VL media tokens do not match the provided media after t

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/d9ae57456e60bfb4. Report an issue: GitHub.