hiyouga/LlamaFactory · error · ValueError

MOSS-VL encountered a video end token without a matching sta

Error message

MOSS-VL encountered a video end token without a matching start token after tokenization. Please increase `cutoff_len` if a video placeholder was truncated.

What it means

MOSS-VL post-tokenization check: a vision_end token id appears in input_ids without a preceding vision_start token. This asymmetric state arises when truncation removes the start token of a media block but leaves its end token, corrupting the media-order reconstruction.

Source

Thrown at src/llamafactory/data/mm_plugin.py:582

    ) -> list[str]:
        media_order = []
        video_frame_counts = []
        in_video = False
        current_video_frames = 0
        for token_id in input_ids:
            if token_id == processor.vision_start_token_id:
                if in_video:
                    raise ValueError(
                        "MOSS-VL encountered nested video token blocks after tokenization. "
                        "Please increase `cutoff_len` if a video placeholder was truncated."
                    )

                media_order.append("video")
                in_video = True
                current_video_frames = 0
            elif token_id == processor.vision_end_token_id:
                if not in_video:
                    raise ValueError(
                        "MOSS-VL encountered a video end token without a matching start token after tokenization. "
                        "Please increase `cutoff_len` if a video placeholder was truncated."
                    )

                video_frame_counts.append(current_video_frames)
                in_video = False
            elif token_id == processor.image_token_id:
                if in_video:
                    current_video_frames += 1
                else:
                    media_order.append("image")

        if in_video:
            raise ValueError(
                "MOSS-VL encountered an incomplete video token block after tokenization. "
                "Please increase `cutoff_len` or reduce `video_maxlen`."
            )

View on GitHub (pinned to f28afaf635)

Solutions

Increase cutoff_len so no visual placeholder is truncated.
Reduce media token budget: fewer/smaller images, lower video_maxlen.
Inspect the offending sample (enable data debugging logs) and cap its media count.

Example fix

### before
cutoff_len: 4096  # too small for the video placeholder
### after
cutoff_len: 16384

Defensive patterns

Strategy: validation

Validate before calling

ids = tokenizer(text).input_ids
starts = ids.count(processor.vision_start_token_id)
ends = ids.count(processor.vision_end_token_id)
assert starts == ends, f'media block mismatch ({starts} vs {ends}); raise cutoff_len'

Prevention

Treat any 'after tokenization' MOSS-VL error as a truncation symptom; fix cutoff_len first.
Log per-sample token lengths during preprocessing to catch outliers early.

When it happens

Trigger: Same root cause as the nested-block error: cutoff_len truncates a MOSS-VL sample so that a media block's start token is cut while its end token (or a later block's end token) survives in the tokenized sequence.

Common situations: cutoff_len too small relative to expanded image/video placeholders; packing/tokenization settings that drop leading tokens of long multimodal samples.

Related errors

AI-assisted analysis of hiyouga/LlamaFactory@f28afaf635 (2026-08-14). Data as JSON: /api/errors/1d6079d34d1b933a. Report an issue: GitHub.