sgl-project/sglang · error · ValueError

The number of image placeholders does not match…

Error message

The number of image placeholders does not match img_grid_thw entries.

What it means

The Kimi placeholder-expansion loop finished but consumed fewer placeholders than the number of grid-derived image token counts: img_idx != len(image_token_counts). This is the mirror of the exceeds case — the prompt has fewer image_token_id tokens than images/grids supplied.

Solutions

1. Ensure exactly one image placeholder token per image in the prompt
2. If images were removed, also remove their grid entries so counts align
3. Pre-validate: count of image_token_id in input_ids == len(image_token_counts) before calling

Example fix

// before
prompt = f"{IMG} look"  # 1 placeholder
build_from_grids(prompt, grids=[g1, g2])  # 2 images

// after
prompt = f"{IMG} {IMG} look"  # 2 placeholders, 2 images
build_from_grids(prompt, grids=[g1, g2])

n_ph = sum(1 for t in input_ids if t == image_token_id)
assert n_ph == len(image_token_counts), f"{n_ph} placeholders != {len(image_token_counts)} grids"

When it happens

Trigger: Calling get_mm_data / _build_kimi_mm_data_from_grids with a prompt containing fewer image placeholder tokens than img_grid_thw entries (extra images passed but no placeholders for them).

Common situations: Sending multiple images with a template that only emits one placeholder, dropping placeholder tokens during custom prompt assembly, or images appended to image_data without updating the prompt.

Related errors

• Kimi image placeholders must map one-to-one to image data…
• The number of image placeholders exceeds img_grid_thw…
• Expected image placeholder token(s), found .
• Expected one original size for each K3 image.
• Grid dim ( ) not found in

AI-assisted analysis of sgl-project/sglang@0132848349 (2026-08-28). Data as JSON: /api/errors/6103f65cb1eaf3e9. Report an issue: GitHub.

Appendix: source

Thrown at python/sglang/srt/multimodal/processors/kimi_common.py:121

        for token in prompt:
            if token != image_token_id:
                input_ids.append(token)
                continue

            if img_idx >= len(image_token_counts):
                raise ValueError(
                    "The number of image placeholders exceeds img_grid_thw entries."
                )

            num_tokens = image_token_counts[img_idx]
            start = len(input_ids)
            input_ids.extend([image_token_id] * num_tokens)
            offsets.append((start, len(input_ids) - 1))
            img_idx += 1

        if img_idx != len(image_token_counts):
            raise ValueError(
                "The number of image placeholders does not match img_grid_thw entries."
            )

        image_embeddings = embeddings[Modality.IMAGE]
        mm_items = []
        consumed = 0
        for start, end in offsets:
            num_tokens = end - start + 1
            embedding_slice = image_embeddings[consumed : consumed + num_tokens]
            consumed += num_tokens
            mm_items.append(
                MultimodalDataItem(
                    modality=Modality.IMAGE,
                    offsets=[(start, end)],
                    precomputed_embeddings=embedding_slice,
                )
            )

View on GitHub (pinned to 0132848349)