hiyouga/LlamaFactory · error · ValueError

MOSS-VL media lengths do not consume all provided inputs.

Error message

MOSS-VL media lengths do not consume all provided inputs.

What it means

MOSS-VL _get_mm_inputs slices images/videos per sample using imglen/vidlen; after the loop, image_offset and video_offset must have consumed exactly all provided images and videos. A mismatch means the per-sample lens (from placeholder counts) disagree with the actual media list lengths.

Source

Thrown at src/llamafactory/data/mm_plugin.py:759

                media_nums_per_sample.append(1)
                continue

            image_index = 0
            video_index = 0
            for modality in media_order:
                if modality == "image":
                    final_pixel_values.append(image_chunks[image_index])
                    final_grid_thw.append(image_grids[image_index])
                    image_index += 1
                else:
                    final_pixel_values.append(video_chunks[video_index])
                    final_grid_thw.append(video_grids[video_index])
                    video_index += 1

            media_nums_per_sample.append(len(media_order))

        if image_offset != len(images) or video_offset != len(videos):
            raise ValueError("MOSS-VL media lengths do not consume all provided inputs.")

        mm_inputs = {
            "pixel_values": torch.cat(final_pixel_values, dim=0),
            "grid_thw": torch.stack(final_grid_thw),
            "media_nums_per_sample": media_nums_per_sample,
        }
        mm_inputs["cross_attention_mask"] = self._create_cross_attention_mask(
            batch_ids,
            mm_inputs["grid_thw"],
            media_nums_per_sample,
            processor.image_token_id,
            padding_side=processor.tokenizer.padding_side,
        )
        return mm_inputs

    def post_process_mossvl_inputs(
        self,
        features: dict[str, "torch.Tensor"],

View on GitHub (pinned to f28afaf635)

Solutions

1. Increase cutoff_len so no placeholders are truncated (root cause in most cases).
2. Ensure every sample's placeholder counts equal its media list lengths.
3. If calling the API manually, precheck sum(imglens) == len(images) and sum(vidlens) == len(videos).

Example fix

### before
cutoff_len: 4096  # placeholders truncated -> lens sum < media count
### after
cutoff_len: 16384

assert sum(imglens) == len(images) and sum(vidlens) == len(videos), 'media lens do not consume all inputs (truncation?)'

When it happens

Trigger: imglens summing to less than len(images), or vidlens summing to less than len(videos) — e.g. truncated placeholders lowering per-sample counts while the raw media list still holds all items.

Common situations: cutoff_len truncation removing media placeholders (counts drop) while media lists stay full; inconsistent dataset rows where media counts differ from placeholder counts across samples.

Related errors

• MOSS-VL media tokens do not match the provided media after t
• 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 video frame tokens do not match the processed video

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/35e956f1e72c9e8f. Report an issue: GitHub.